diff --git a/README.md b/README.md
index 6a68f268a8b1..a542ed4d6f6f 100644
--- a/README.md
+++ b/README.md
@@ -12,6 +12,7 @@ This ui will let you design and execute advanced stable diffusion pipelines usin
 ## Features
 - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything.
 - Fully supports SD1.x, SD2.x, [SDXL](https://comfyanonymous.github.io/ComfyUI_examples/sdxl/), [Stable Video Diffusion](https://comfyanonymous.github.io/ComfyUI_examples/video/), [Stable Cascade](https://comfyanonymous.github.io/ComfyUI_examples/stable_cascade/), [SD3](https://comfyanonymous.github.io/ComfyUI_examples/sd3/) and [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
+- [Flux](https://comfyanonymous.github.io/ComfyUI_examples/flux/)
 - Asynchronous Queue system
 - Many optimizations: Only re-executes the parts of the workflow that changes between executions.
 - Smart memory management: can automatically run models on GPUs with as low as 1GB vram.
diff --git a/comfy/clip_model.py b/comfy/clip_model.py
index ab775309cc7f..3c67b737aada 100644
--- a/comfy/clip_model.py
+++ b/comfy/clip_model.py
@@ -1,5 +1,6 @@
 import torch
 from comfy.ldm.modules.attention import optimized_attention_for_device
+import comfy.ops
 
 class CLIPAttention(torch.nn.Module):
     def __init__(self, embed_dim, heads, dtype, device, operations):
@@ -71,13 +72,13 @@ def forward(self, x, mask=None, intermediate_output=None):
         return x, intermediate
 
 class CLIPEmbeddings(torch.nn.Module):
-    def __init__(self, embed_dim, vocab_size=49408, num_positions=77, dtype=None, device=None):
+    def __init__(self, embed_dim, vocab_size=49408, num_positions=77, dtype=None, device=None, operations=None):
         super().__init__()
-        self.token_embedding = torch.nn.Embedding(vocab_size, embed_dim, dtype=dtype, device=device)
-        self.position_embedding = torch.nn.Embedding(num_positions, embed_dim, dtype=dtype, device=device)
+        self.token_embedding = operations.Embedding(vocab_size, embed_dim, dtype=dtype, device=device)
+        self.position_embedding = operations.Embedding(num_positions, embed_dim, dtype=dtype, device=device)
 
-    def forward(self, input_tokens):
-        return self.token_embedding(input_tokens) + self.position_embedding.weight
+    def forward(self, input_tokens, dtype=torch.float32):
+        return self.token_embedding(input_tokens, out_dtype=dtype) + comfy.ops.cast_to(self.position_embedding.weight, dtype=dtype, device=input_tokens.device)
 
 
 class CLIPTextModel_(torch.nn.Module):
@@ -90,12 +91,12 @@ def __init__(self, config_dict, dtype, device, operations):
         self.eos_token_id = config_dict["eos_token_id"]
 
         super().__init__()
-        self.embeddings = CLIPEmbeddings(embed_dim, dtype=torch.float32, device=device)
+        self.embeddings = CLIPEmbeddings(embed_dim, dtype=dtype, device=device, operations=operations)
         self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
         self.final_layer_norm = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
 
-    def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True):
-        x = self.embeddings(input_tokens)
+    def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=torch.float32):
+        x = self.embeddings(input_tokens, dtype=dtype)
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
@@ -154,11 +155,11 @@ def __init__(self, embed_dim, num_channels=3, patch_size=14, image_size=224, dty
 
         num_patches = (image_size // patch_size) ** 2
         num_positions = num_patches + 1
-        self.position_embedding = torch.nn.Embedding(num_positions, embed_dim, dtype=dtype, device=device)
+        self.position_embedding = operations.Embedding(num_positions, embed_dim, dtype=dtype, device=device)
 
     def forward(self, pixel_values):
         embeds = self.patch_embedding(pixel_values).flatten(2).transpose(1, 2)
-        return torch.cat([self.class_embedding.to(embeds.device).expand(pixel_values.shape[0], 1, -1), embeds], dim=1) + self.position_embedding.weight.to(embeds.device)
+        return torch.cat([comfy.ops.cast_to_input(self.class_embedding, embeds).expand(pixel_values.shape[0], 1, -1), embeds], dim=1) + comfy.ops.cast_to_input(self.position_embedding.weight, embeds)
 
 
 class CLIPVision(torch.nn.Module):
@@ -170,7 +171,7 @@ def __init__(self, config_dict, dtype, device, operations):
         intermediate_size = config_dict["intermediate_size"]
         intermediate_activation = config_dict["hidden_act"]
 
-        self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], dtype=torch.float32, device=device, operations=operations)
+        self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], dtype=dtype, device=device, operations=operations)
         self.pre_layrnorm = operations.LayerNorm(embed_dim)
         self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
         self.post_layernorm = operations.LayerNorm(embed_dim)
diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py
index 4b4a9eda2ca5..34c7bb3da22a 100644
--- a/comfy/latent_formats.py
+++ b/comfy/latent_formats.py
@@ -139,3 +139,14 @@ def process_out(self, latent):
 
 class StableAudio1(LatentFormat):
     latent_channels = 64
+
+class Flux(SD3):
+    def __init__(self):
+        self.scale_factor = 0.3611
+        self.shift_factor = 0.1159
+
+    def process_in(self, latent):
+        return (latent - self.shift_factor) * self.scale_factor
+
+    def process_out(self, latent):
+        return (latent / self.scale_factor) + self.shift_factor
diff --git a/comfy/ldm/flux/layers.py b/comfy/ldm/flux/layers.py
new file mode 100644
index 000000000000..99f498106ebd
--- /dev/null
+++ b/comfy/ldm/flux/layers.py
@@ -0,0 +1,256 @@
+import math
+from dataclasses import dataclass
+
+import torch
+from einops import rearrange
+from torch import Tensor, nn
+
+from .math import attention, rope
+import comfy.ops
+
+class EmbedND(nn.Module):
+    def __init__(self, dim: int, theta: int, axes_dim: list):
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+        self.axes_dim = axes_dim
+
+    def forward(self, ids: Tensor) -> Tensor:
+        n_axes = ids.shape[-1]
+        emb = torch.cat(
+            [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
+            dim=-3,
+        )
+
+        return emb.unsqueeze(1)
+
+
+def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 1000.0):
+    """
+    Create sinusoidal timestep embeddings.
+    :param t: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an (N, D) Tensor of positional embeddings.
+    """
+    t = time_factor * t
+    half = dim // 2
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
+        t.device
+    )
+
+    args = t[:, None].float() * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    if torch.is_floating_point(t):
+        embedding = embedding.to(t)
+    return embedding
+
+
+class MLPEmbedder(nn.Module):
+    def __init__(self, in_dim: int, hidden_dim: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.in_layer = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+        self.silu = nn.SiLU()
+        self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.out_layer(self.silu(self.in_layer(x)))
+
+
+class RMSNorm(torch.nn.Module):
+    def __init__(self, dim: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.scale = nn.Parameter(torch.empty((dim), dtype=dtype, device=device))
+
+    def forward(self, x: Tensor):
+        x_dtype = x.dtype
+        x = x.float()
+        rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + 1e-6)
+        return (x * rrms).to(dtype=x_dtype) * comfy.ops.cast_to(self.scale, dtype=x_dtype, device=x.device)
+
+
+class QKNorm(torch.nn.Module):
+    def __init__(self, dim: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.query_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
+        self.key_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple:
+        q = self.query_norm(q)
+        k = self.key_norm(k)
+        return q.to(v), k.to(v)
+
+
+class SelfAttention(nn.Module):
+    def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+
+        self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
+        self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
+        self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
+
+    def forward(self, x: Tensor, pe: Tensor) -> Tensor:
+        qkv = self.qkv(x)
+        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        q, k = self.norm(q, k, v)
+        x = attention(q, k, v, pe=pe)
+        x = self.proj(x)
+        return x
+
+
+@dataclass
+class ModulationOut:
+    shift: Tensor
+    scale: Tensor
+    gate: Tensor
+
+
+class Modulation(nn.Module):
+    def __init__(self, dim: int, double: bool, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.is_double = double
+        self.multiplier = 6 if double else 3
+        self.lin = operations.Linear(dim, self.multiplier * dim, bias=True, dtype=dtype, device=device)
+
+    def forward(self, vec: Tensor) -> tuple:
+        out = self.lin(nn.functional.silu(vec))[:, None, :].chunk(self.multiplier, dim=-1)
+
+        return (
+            ModulationOut(*out[:3]),
+            ModulationOut(*out[3:]) if self.is_double else None,
+        )
+
+
+class DoubleStreamBlock(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
+        self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+
+        self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.img_mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.GELU(approximate="tanh"),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+        self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
+        self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+
+        self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.txt_mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.GELU(approximate="tanh"),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+    def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor):
+        img_mod1, img_mod2 = self.img_mod(vec)
+        txt_mod1, txt_mod2 = self.txt_mod(vec)
+
+        # prepare image for attention
+        img_modulated = self.img_norm1(img)
+        img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
+        img_qkv = self.img_attn.qkv(img_modulated)
+        img_q, img_k, img_v = rearrange(img_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
+
+        # prepare txt for attention
+        txt_modulated = self.txt_norm1(txt)
+        txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
+        txt_qkv = self.txt_attn.qkv(txt_modulated)
+        txt_q, txt_k, txt_v = rearrange(txt_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
+
+        # run actual attention
+        q = torch.cat((txt_q, img_q), dim=2)
+        k = torch.cat((txt_k, img_k), dim=2)
+        v = torch.cat((txt_v, img_v), dim=2)
+
+        attn = attention(q, k, v, pe=pe)
+        txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
+
+        # calculate the img bloks
+        img = img + img_mod1.gate * self.img_attn.proj(img_attn)
+        img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift)
+
+        # calculate the txt bloks
+        txt = txt + txt_mod1.gate * self.txt_attn.proj(txt_attn)
+        txt = txt + txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
+        return img, txt
+
+
+class SingleStreamBlock(nn.Module):
+    """
+    A DiT block with parallel linear layers as described in
+    https://arxiv.org/abs/2302.05442 and adapted modulation interface.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qk_scale: float = None,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.hidden_dim = hidden_size
+        self.num_heads = num_heads
+        head_dim = hidden_size // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        # qkv and mlp_in
+        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device)
+        # proj and mlp_out
+        self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device)
+
+        self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
+
+        self.hidden_size = hidden_size
+        self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+
+        self.mlp_act = nn.GELU(approximate="tanh")
+        self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, x: Tensor, vec: Tensor, pe: Tensor) -> Tensor:
+        mod, _ = self.modulation(vec)
+        x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift
+        qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+
+        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        q, k = self.norm(q, k, v)
+
+        # compute attention
+        attn = attention(q, k, v, pe=pe)
+        # compute activation in mlp stream, cat again and run second linear layer
+        output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
+        return x + mod.gate * output
+
+
+class LastLayer(nn.Module):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device)
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device))
+
+    def forward(self, x: Tensor, vec: Tensor) -> Tensor:
+        shift, scale = self.adaLN_modulation(vec).chunk(2, dim=1)
+        x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
+        x = self.linear(x)
+        return x
diff --git a/comfy/ldm/flux/math.py b/comfy/ldm/flux/math.py
new file mode 100644
index 000000000000..88c2b6bb409c
--- /dev/null
+++ b/comfy/ldm/flux/math.py
@@ -0,0 +1,35 @@
+import torch
+from einops import rearrange
+from torch import Tensor
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.model_management
+
+def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor) -> Tensor:
+    q, k = apply_rope(q, k, pe)
+
+    heads = q.shape[1]
+    x = optimized_attention(q, k, v, heads, skip_reshape=True)
+    return x
+
+
+def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
+    assert dim % 2 == 0
+    if comfy.model_management.is_device_mps(pos.device):
+        device = torch.device("cpu")
+    else:
+        device = pos.device
+
+    scale = torch.linspace(0, (dim - 2) / dim, steps=dim//2, dtype=torch.float64, device=device)
+    omega = 1.0 / (theta**scale)
+    out = torch.einsum("...n,d->...nd", pos.to(dtype=torch.float32, device=device), omega)
+    out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1)
+    out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
+    return out.to(dtype=torch.float32, device=pos.device)
+
+
+def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
+    xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)
+    xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2)
+    xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
+    xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
+    return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)
diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.py
new file mode 100644
index 000000000000..21f1a877edea
--- /dev/null
+++ b/comfy/ldm/flux/model.py
@@ -0,0 +1,138 @@
+#Original code can be found on: https://github.com/black-forest-labs/flux
+
+from dataclasses import dataclass
+
+import torch
+from torch import Tensor, nn
+
+from .layers import (
+    DoubleStreamBlock,
+    EmbedND,
+    LastLayer,
+    MLPEmbedder,
+    SingleStreamBlock,
+    timestep_embedding,
+)
+
+from einops import rearrange, repeat
+
+@dataclass
+class FluxParams:
+    in_channels: int
+    vec_in_dim: int
+    context_in_dim: int
+    hidden_size: int
+    mlp_ratio: float
+    num_heads: int
+    depth: int
+    depth_single_blocks: int
+    axes_dim: list
+    theta: int
+    qkv_bias: bool
+    guidance_embed: bool
+
+
+class Flux(nn.Module):
+    """
+    Transformer model for flow matching on sequences.
+    """
+
+    def __init__(self, image_model=None, dtype=None, device=None, operations=None, **kwargs):
+        super().__init__()
+        self.dtype = dtype
+        params = FluxParams(**kwargs)
+        self.params = params
+        self.in_channels = params.in_channels
+        self.out_channels = self.in_channels
+        if params.hidden_size % params.num_heads != 0:
+            raise ValueError(
+                f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
+            )
+        pe_dim = params.hidden_size // params.num_heads
+        if sum(params.axes_dim) != pe_dim:
+            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
+        self.hidden_size = params.hidden_size
+        self.num_heads = params.num_heads
+        self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
+        self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations)
+        self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
+        self.guidance_in = (
+            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity()
+        )
+        self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                    qkv_bias=params.qkv_bias,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(params.depth)
+            ]
+        )
+
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
+                for _ in range(params.depth_single_blocks)
+            ]
+        )
+
+        self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        y: Tensor,
+        guidance: Tensor = None,
+    ) -> Tensor:
+        if img.ndim != 3 or txt.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
+        # running on sequences img
+        img = self.img_in(img)
+        vec = self.time_in(timestep_embedding(timesteps, 256).to(img.dtype))
+        if self.params.guidance_embed:
+            if guidance is None:
+                raise ValueError("Didn't get guidance strength for guidance distilled model.")
+            vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))
+
+        vec = vec + self.vector_in(y)
+        txt = self.txt_in(txt)
+
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        for block in self.double_blocks:
+            img, txt = block(img=img, txt=txt, vec=vec, pe=pe)
+
+        img = torch.cat((txt, img), 1)
+        for block in self.single_blocks:
+            img = block(img, vec=vec, pe=pe)
+        img = img[:, txt.shape[1] :, ...]
+
+        img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
+        return img
+
+    def forward(self, x, timestep, context, y, guidance, **kwargs):
+        bs, c, h, w = x.shape
+        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+
+        h_len = (h // 2)
+        w_len = (w // 2)
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[..., 1] = img_ids[..., 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype)[:, None]
+        img_ids[..., 2] = img_ids[..., 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype)[None, :]
+        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+
+        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance)
+        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)
diff --git a/comfy/ldm/hydit/models.py b/comfy/ldm/hydit/models.py
index 37836fc312bd..c70dbf92a9ed 100644
--- a/comfy/ldm/hydit/models.py
+++ b/comfy/ldm/hydit/models.py
@@ -355,7 +355,7 @@ def forward(self,
         if self.use_style_cond:
             if style is None:
                 style = torch.zeros((extra_vec.shape[0],), device=x.device, dtype=torch.int)
-            style_embedding = self.style_embedder(style)
+            style_embedding = self.style_embedder(style, out_dtype=x.dtype)
             extra_vec = torch.cat([extra_vec, style_embedding], dim=1)
 
         # Concatenate all extra vectors
diff --git a/comfy/model_base.py b/comfy/model_base.py
index 9d60c1fb4e61..7c7b4c3ff651 100644
--- a/comfy/model_base.py
+++ b/comfy/model_base.py
@@ -10,6 +10,8 @@
 import comfy.ldm.hydit.models
 import comfy.ldm.audio.dit
 import comfy.ldm.audio.embedders
+import comfy.ldm.flux.model
+
 import comfy.model_management
 import comfy.conds
 import comfy.ops
@@ -26,6 +28,7 @@ class ModelType(Enum):
     EDM = 5
     FLOW = 6
     V_PREDICTION_CONTINUOUS = 7
+    FLUX = 8
 
 
 from comfy.model_sampling import EPS, V_PREDICTION, EDM, ModelSamplingDiscrete, ModelSamplingContinuousEDM, StableCascadeSampling, ModelSamplingContinuousV
@@ -53,6 +56,9 @@ def model_sampling(model_config, model_type):
     elif model_type == ModelType.V_PREDICTION_CONTINUOUS:
         c = V_PREDICTION
         s = ModelSamplingContinuousV
+    elif model_type == ModelType.FLUX:
+        c = comfy.model_sampling.CONST
+        s = comfy.model_sampling.ModelSamplingFlux
 
     class ModelSampling(s, c):
         pass
@@ -681,3 +687,18 @@ def extra_conds(self, **kwargs):
 
         out['image_meta_size'] = comfy.conds.CONDRegular(torch.FloatTensor([[height, width, target_height, target_width, 0, 0]]))
         return out
+
+class Flux(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.flux.model.Flux)
+
+    def encode_adm(self, **kwargs):
+        return kwargs["pooled_output"]
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([kwargs.get("guidance", 3.5)]))
+        return out
diff --git a/comfy/model_detection.py b/comfy/model_detection.py
index ea4959370405..dda9797b7b92 100644
--- a/comfy/model_detection.py
+++ b/comfy/model_detection.py
@@ -128,6 +128,23 @@ def detect_unet_config(state_dict, key_prefix):
             unet_config["image_model"] = "hydit1"
         return unet_config
 
+    if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys: #Flux
+        dit_config = {}
+        dit_config["image_model"] = "flux"
+        dit_config["in_channels"] = 64
+        dit_config["vec_in_dim"] = 768
+        dit_config["context_in_dim"] = 4096
+        dit_config["hidden_size"] = 3072
+        dit_config["mlp_ratio"] = 4.0
+        dit_config["num_heads"] = 24
+        dit_config["depth"] = 19
+        dit_config["depth_single_blocks"] = 38
+        dit_config["axes_dim"] = [16, 56, 56]
+        dit_config["theta"] = 10000
+        dit_config["qkv_bias"] = True
+        dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 
diff --git a/comfy/model_management.py b/comfy/model_management.py
index 047193290fa2..de4bd4426ac8 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -318,7 +318,7 @@ def __eq__(self, other):
         return self.model is other.model
 
 def minimum_inference_memory():
-    return (1024 * 1024 * 1024)
+    return (1024 * 1024 * 1024) * 1.2
 
 def unload_model_clones(model, unload_weights_only=True, force_unload=True):
     to_unload = []
@@ -446,7 +446,7 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False):
         if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM):
             model_size = loaded_model.model_memory_required(torch_dev)
             current_free_mem = get_free_memory(torch_dev)
-            lowvram_model_memory = int(max(64 * (1024 * 1024), (current_free_mem - 1024 * (1024 * 1024)) / 1.3 ))
+            lowvram_model_memory = int(max(64 * (1024 * 1024), (current_free_mem - extra_mem)))
             if model_size <= (current_free_mem - inference_memory): #only switch to lowvram if really necessary
                 lowvram_model_memory = 0
 
@@ -649,18 +649,29 @@ def supports_cast(device, dtype): #TODO
         return True
     if dtype == torch.float16:
         return True
-    if is_device_mps(device):
-        return False
     if directml_enabled: #TODO: test this
         return False
     if dtype == torch.bfloat16:
         return True
+    if is_device_mps(device):
+        return False
     if dtype == torch.float8_e4m3fn:
         return True
     if dtype == torch.float8_e5m2:
         return True
     return False
 
+def pick_weight_dtype(dtype, fallback_dtype, device=None):
+    if dtype is None:
+        dtype = fallback_dtype
+    elif dtype_size(dtype) > dtype_size(fallback_dtype):
+        dtype = fallback_dtype
+
+    if not supports_cast(device, dtype):
+        dtype = fallback_dtype
+
+    return dtype
+
 def device_supports_non_blocking(device):
     if is_device_mps(device):
         return False #pytorch bug? mps doesn't support non blocking
@@ -876,9 +887,9 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
         if is_device_cpu(device): #TODO ? bf16 works on CPU but is extremely slow
             return False
 
-    if device is not None: #TODO not sure about mps bf16 support
+    if device is not None:
         if is_device_mps(device):
-            return False
+            return True
 
     if FORCE_FP32:
         return False
diff --git a/comfy/model_sampling.py b/comfy/model_sampling.py
index 25bb7e043b62..4a0f2db60723 100644
--- a/comfy/model_sampling.py
+++ b/comfy/model_sampling.py
@@ -272,3 +272,43 @@ def percent_to_sigma(self, percent):
 
         percent = 1.0 - percent
         return self.sigma(torch.tensor(percent))
+
+
+def flux_time_shift(mu: float, sigma: float, t):
+    return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
+
+class ModelSamplingFlux(torch.nn.Module):
+    def __init__(self, model_config=None):
+        super().__init__()
+        if model_config is not None:
+            sampling_settings = model_config.sampling_settings
+        else:
+            sampling_settings = {}
+
+        self.set_parameters(shift=sampling_settings.get("shift", 1.15))
+
+    def set_parameters(self, shift=1.15, timesteps=10000):
+        self.shift = shift
+        ts = self.sigma((torch.arange(1, timesteps + 1, 1) / timesteps))
+        self.register_buffer('sigmas', ts)
+
+    @property
+    def sigma_min(self):
+        return self.sigmas[0]
+
+    @property
+    def sigma_max(self):
+        return self.sigmas[-1]
+
+    def timestep(self, sigma):
+        return sigma
+
+    def sigma(self, timestep):
+        return flux_time_shift(self.shift, 1.0, timestep)
+
+    def percent_to_sigma(self, percent):
+        if percent <= 0.0:
+            return 1.0
+        if percent >= 1.0:
+            return 0.0
+        return 1.0 - percent
diff --git a/comfy/ops.py b/comfy/ops.py
index 6d88a1b9b7d5..47e8d7a9d5d6 100644
--- a/comfy/ops.py
+++ b/comfy/ops.py
@@ -19,17 +19,27 @@
 import torch
 import comfy.model_management
 
+
+def cast_to(weight, dtype=None, device=None, non_blocking=False):
+    return weight.to(device=device, dtype=dtype, non_blocking=non_blocking)
+
 def cast_to_input(weight, input, non_blocking=False):
-    return weight.to(device=input.device, dtype=input.dtype, non_blocking=non_blocking)
+    return cast_to(weight, input.dtype, input.device, non_blocking=non_blocking)
+
+def cast_bias_weight(s, input=None, dtype=None, device=None):
+    if input is not None:
+        if dtype is None:
+            dtype = input.dtype
+        if device is None:
+            device = input.device
 
-def cast_bias_weight(s, input):
     bias = None
-    non_blocking = comfy.model_management.device_should_use_non_blocking(input.device)
+    non_blocking = comfy.model_management.device_should_use_non_blocking(device)
     if s.bias is not None:
-        bias = cast_to_input(s.bias, input, non_blocking=non_blocking)
+        bias = cast_to(s.bias, dtype, device, non_blocking=non_blocking)
         if s.bias_function is not None:
             bias = s.bias_function(bias)
-    weight = cast_to_input(s.weight, input, non_blocking=non_blocking)
+    weight = cast_to(s.weight, dtype, device, non_blocking=non_blocking)
     if s.weight_function is not None:
         weight = s.weight_function(weight)
     return weight, bias
@@ -176,14 +186,19 @@ def reset_parameters(self):
             self.bias = None
             return None
 
-        def forward_comfy_cast_weights(self, input):
-            weight, bias = cast_bias_weight(self, input)
-            return torch.nn.functional.embedding(input, weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse)
+        def forward_comfy_cast_weights(self, input, out_dtype=None):
+            output_dtype = out_dtype
+            if self.weight.dtype == torch.float16 or self.weight.dtype == torch.bfloat16:
+                out_dtype = None
+            weight, bias = cast_bias_weight(self, device=input.device, dtype=out_dtype)
+            return torch.nn.functional.embedding(input, weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse).to(dtype=output_dtype)
 
         def forward(self, *args, **kwargs):
             if self.comfy_cast_weights:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
+                if "out_dtype" in kwargs:
+                    kwargs.pop("out_dtype")
                 return super().forward(*args, **kwargs)
 
     @classmethod
diff --git a/comfy/sd.py b/comfy/sd.py
index 8bf8d1087dea..41ce18c803fa 100644
--- a/comfy/sd.py
+++ b/comfy/sd.py
@@ -23,6 +23,7 @@
 import comfy.text_encoders.sa_t5
 import comfy.text_encoders.aura_t5
 import comfy.text_encoders.hydit
+import comfy.text_encoders.flux
 
 import comfy.model_patcher
 import comfy.lora
@@ -387,6 +388,7 @@ class CLIPType(Enum):
     SD3 = 3
     STABLE_AUDIO = 4
     HUNYUAN_DIT = 5
+    FLUX = 6
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION):
     clip_data = []
@@ -438,6 +440,15 @@ class EmptyClass:
         elif clip_type == CLIPType.HUNYUAN_DIT:
             clip_target.clip = comfy.text_encoders.hydit.HyditModel
             clip_target.tokenizer = comfy.text_encoders.hydit.HyditTokenizer
+        elif clip_type == CLIPType.FLUX:
+            weight_name = "encoder.block.23.layer.1.DenseReluDense.wi_1.weight"
+            weight = clip_data[0].get(weight_name, clip_data[1].get(weight_name, None))
+            dtype_t5 = None
+            if weight is not None:
+                dtype_t5 = weight.dtype
+
+            clip_target.clip = comfy.text_encoders.flux.flux_clip(dtype_t5=dtype_t5)
+            clip_target.tokenizer = comfy.text_encoders.flux.FluxTokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer
@@ -556,7 +567,7 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
     return (model_patcher, clip, vae, clipvision)
 
 
-def load_unet_state_dict(sd): #load unet in diffusers or regular format
+def load_unet_state_dict(sd, dtype=None): #load unet in diffusers or regular format
 
     #Allow loading unets from checkpoint files
     diffusion_model_prefix = model_detection.unet_prefix_from_state_dict(sd)
@@ -565,7 +576,6 @@ def load_unet_state_dict(sd): #load unet in diffusers or regular format
         sd = temp_sd
 
     parameters = comfy.utils.calculate_parameters(sd)
-    unet_dtype = model_management.unet_dtype(model_params=parameters)
     load_device = model_management.get_torch_device()
     model_config = model_detection.model_config_from_unet(sd, "")
 
@@ -592,7 +602,11 @@ def load_unet_state_dict(sd): #load unet in diffusers or regular format
                     logging.warning("{} {}".format(diffusers_keys[k], k))
 
     offload_device = model_management.unet_offload_device()
-    unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=model_config.supported_inference_dtypes)
+    if dtype is None:
+        unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=model_config.supported_inference_dtypes)
+    else:
+        unet_dtype = dtype
+
     manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)
     model_config.set_inference_dtype(unet_dtype, manual_cast_dtype)
     model = model_config.get_model(new_sd, "")
@@ -603,9 +617,9 @@ def load_unet_state_dict(sd): #load unet in diffusers or regular format
         logging.info("left over keys in unet: {}".format(left_over))
     return comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=offload_device)
 
-def load_unet(unet_path):
+def load_unet(unet_path, dtype=None):
     sd = comfy.utils.load_torch_file(unet_path)
-    model = load_unet_state_dict(sd)
+    model = load_unet_state_dict(sd, dtype=dtype)
     if model is None:
         logging.error("ERROR UNSUPPORTED UNET {}".format(unet_path))
         raise RuntimeError("ERROR: Could not detect model type of: {}".format(unet_path))
diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py
index f209bed41ba7..d32121d1b219 100644
--- a/comfy/sd1_clip.py
+++ b/comfy/sd1_clip.py
@@ -94,7 +94,8 @@ def __init__(self, version="openai/clip-vit-large-patch14", device="cpu", max_le
         with open(textmodel_json_config) as f:
             config = json.load(f)
 
-        self.transformer = model_class(config, dtype, device, comfy.ops.manual_cast)
+        self.operations = comfy.ops.manual_cast
+        self.transformer = model_class(config, dtype, device, self.operations)
         self.num_layers = self.transformer.num_layers
 
         self.max_length = max_length
@@ -161,7 +162,7 @@ def set_up_textual_embeddings(self, tokens, current_embeds):
 
         n = token_dict_size
         if len(embedding_weights) > 0:
-            new_embedding = torch.nn.Embedding(next_new_token + 1, current_embeds.weight.shape[1], device=current_embeds.weight.device, dtype=current_embeds.weight.dtype)
+            new_embedding = self.operations.Embedding(next_new_token + 1, current_embeds.weight.shape[1], device=current_embeds.weight.device, dtype=current_embeds.weight.dtype)
             new_embedding.weight[:token_dict_size] = current_embeds.weight
             for x in embedding_weights:
                 new_embedding.weight[n] = x
@@ -194,7 +195,7 @@ def forward(self, tokens):
         if self.enable_attention_masks:
             attention_mask_model = attention_mask
 
-        outputs = self.transformer(tokens, attention_mask_model, intermediate_output=self.layer_idx, final_layer_norm_intermediate=self.layer_norm_hidden_state)
+        outputs = self.transformer(tokens, attention_mask_model, intermediate_output=self.layer_idx, final_layer_norm_intermediate=self.layer_norm_hidden_state, dtype=torch.float32)
         self.transformer.set_input_embeddings(backup_embeds)
 
         if self.layer == "last":
diff --git a/comfy/supported_models.py b/comfy/supported_models.py
index ddd0c173bbea..43e8f5d1b157 100644
--- a/comfy/supported_models.py
+++ b/comfy/supported_models.py
@@ -9,6 +9,7 @@
 import comfy.text_encoders.sa_t5
 import comfy.text_encoders.aura_t5
 import comfy.text_encoders.hydit
+import comfy.text_encoders.flux
 
 from . import supported_models_base
 from . import latent_formats
@@ -619,7 +620,45 @@ class HunyuanDiT1(HunyuanDiT):
         "linear_end" : 0.03,
     }
 
+class Flux(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "flux",
+        "guidance_embed": True,
+    }
+
+    sampling_settings = {
+    }
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Flux
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Flux(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.FluxClipModel)
+
+class FluxSchnell(Flux):
+    unet_config = {
+        "image_model": "flux",
+        "guidance_embed": False,
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 1.0,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Flux(self, model_type=model_base.ModelType.FLOW, device=device)
+        return out
+
 
-models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, HunyuanDiT, HunyuanDiT1]
+models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, HunyuanDiT, HunyuanDiT1, Flux, FluxSchnell]
 
 models += [SVD_img2vid]
diff --git a/comfy/text_encoders/bert.py b/comfy/text_encoders/bert.py
index b76e76660d16..fc9bac1d24c0 100644
--- a/comfy/text_encoders/bert.py
+++ b/comfy/text_encoders/bert.py
@@ -1,5 +1,6 @@
 import torch
 from comfy.ldm.modules.attention import optimized_attention_for_device
+import comfy.ops
 
 class BertAttention(torch.nn.Module):
     def __init__(self, embed_dim, heads, dtype, device, operations):
@@ -86,19 +87,19 @@ def forward(self, x, mask=None, intermediate_output=None):
 class BertEmbeddings(torch.nn.Module):
     def __init__(self, vocab_size, max_position_embeddings, type_vocab_size, pad_token_id, embed_dim, layer_norm_eps, dtype, device, operations):
         super().__init__()
-        self.word_embeddings = torch.nn.Embedding(vocab_size, embed_dim, padding_idx=pad_token_id, dtype=dtype, device=device)
-        self.position_embeddings = torch.nn.Embedding(max_position_embeddings, embed_dim, dtype=dtype, device=device)
-        self.token_type_embeddings = torch.nn.Embedding(type_vocab_size, embed_dim, dtype=dtype, device=device)
+        self.word_embeddings = operations.Embedding(vocab_size, embed_dim, padding_idx=pad_token_id, dtype=dtype, device=device)
+        self.position_embeddings = operations.Embedding(max_position_embeddings, embed_dim, dtype=dtype, device=device)
+        self.token_type_embeddings = operations.Embedding(type_vocab_size, embed_dim, dtype=dtype, device=device)
 
         self.LayerNorm = operations.LayerNorm(embed_dim, eps=layer_norm_eps, dtype=dtype, device=device)
 
-    def forward(self, input_tokens, token_type_ids=None):
-        x = self.word_embeddings(input_tokens)
-        x += self.position_embeddings.weight[:x.shape[1]]
+    def forward(self, input_tokens, token_type_ids=None, dtype=None):
+        x = self.word_embeddings(input_tokens, out_dtype=dtype)
+        x += comfy.ops.cast_to_input(self.position_embeddings.weight[:x.shape[1]], x)
         if token_type_ids is not None:
-            x += self.token_type_embeddings(token_type_ids)
+            x += self.token_type_embeddings(token_type_ids, out_dtype=x.dtype)
         else:
-            x += self.token_type_embeddings.weight[0]
+            x += comfy.ops.cast_to_input(self.token_type_embeddings.weight[0], x)
         x = self.LayerNorm(x)
         return x
 
@@ -112,8 +113,8 @@ def __init__(self, config_dict, dtype, device, operations):
         self.embeddings = BertEmbeddings(config_dict["vocab_size"], config_dict["max_position_embeddings"], config_dict["type_vocab_size"], config_dict["pad_token_id"], embed_dim, layer_norm_eps, dtype, device, operations)
         self.encoder = BertEncoder(config_dict["num_hidden_layers"], embed_dim, config_dict["intermediate_size"], config_dict["num_attention_heads"], layer_norm_eps, dtype, device, operations)
 
-    def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True):
-        x = self.embeddings(input_tokens)
+    def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
+        x = self.embeddings(input_tokens, dtype=dtype)
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
diff --git a/comfy/text_encoders/flux.py b/comfy/text_encoders/flux.py
new file mode 100644
index 000000000000..849214ce080a
--- /dev/null
+++ b/comfy/text_encoders/flux.py
@@ -0,0 +1,71 @@
+from comfy import sd1_clip
+import comfy.text_encoders.t5
+import comfy.model_management
+from transformers import T5TokenizerFast
+import torch
+import os
+
+class T5XXLModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json")
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5)
+
+class T5XXLTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256)
+
+
+class FluxTokenizer:
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory)
+        self.t5xxl = T5XXLTokenizer(embedding_directory=embedding_directory)
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False):
+        out = {}
+        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids)
+        out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids)
+        return out
+
+    def untokenize(self, token_weight_pair):
+        return self.clip_g.untokenize(token_weight_pair)
+
+    def state_dict(self):
+        return {}
+
+
+class FluxClipModel(torch.nn.Module):
+    def __init__(self, dtype_t5=None, device="cpu", dtype=None):
+        super().__init__()
+        dtype_t5 = comfy.model_management.pick_weight_dtype(dtype_t5, dtype, device)
+        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False)
+        self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5)
+        self.dtypes = set([dtype, dtype_t5])
+
+    def set_clip_options(self, options):
+        self.clip_l.set_clip_options(options)
+        self.t5xxl.set_clip_options(options)
+
+    def reset_clip_options(self):
+        self.clip_l.reset_clip_options()
+        self.t5xxl.reset_clip_options()
+
+    def encode_token_weights(self, token_weight_pairs):
+        token_weight_pairs_l = token_weight_pairs["l"]
+        token_weight_pars_t5 = token_weight_pairs["t5xxl"]
+
+        t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pars_t5)
+        l_out, l_pooled = self.clip_l.encode_token_weights(token_weight_pairs_l)
+        return t5_out, l_pooled
+
+    def load_sd(self, sd):
+        if "text_model.encoder.layers.1.mlp.fc1.weight" in sd:
+            return self.clip_l.load_sd(sd)
+        else:
+            return self.t5xxl.load_sd(sd)
+
+def flux_clip(dtype_t5=None):
+    class FluxClipModel_(FluxClipModel):
+        def __init__(self, device="cpu", dtype=None):
+            super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype)
+    return FluxClipModel_
diff --git a/comfy/text_encoders/hydit.py b/comfy/text_encoders/hydit.py
index ac0c893bcd21..9dfa288b1191 100644
--- a/comfy/text_encoders/hydit.py
+++ b/comfy/text_encoders/hydit.py
@@ -52,8 +52,8 @@ def state_dict(self):
 class HyditModel(torch.nn.Module):
     def __init__(self, device="cpu", dtype=None):
         super().__init__()
-        self.hydit_clip = HyditBertModel()
-        self.mt5xl = MT5XLModel()
+        self.hydit_clip = HyditBertModel(dtype=dtype)
+        self.mt5xl = MT5XLModel(dtype=dtype)
 
         self.dtypes = set()
         if dtype is not None:
diff --git a/comfy/text_encoders/sd3_clip.py b/comfy/text_encoders/sd3_clip.py
index b01fad221a2c..143d884cbe9c 100644
--- a/comfy/text_encoders/sd3_clip.py
+++ b/comfy/text_encoders/sd3_clip.py
@@ -54,14 +54,7 @@ def __init__(self, clip_l=True, clip_g=True, t5=True, dtype_t5=None, device="cpu
             self.clip_g = None
 
         if t5:
-            if dtype_t5 is None:
-                dtype_t5 = dtype
-            elif comfy.model_management.dtype_size(dtype_t5) > comfy.model_management.dtype_size(dtype):
-                dtype_t5 = dtype
-
-            if not comfy.model_management.supports_cast(device, dtype_t5):
-                dtype_t5 = dtype
-
+            dtype_t5 = comfy.model_management.pick_weight_dtype(dtype_t5, dtype, device)
             self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5)
             self.dtypes.add(dtype_t5)
         else:
diff --git a/comfy/text_encoders/t5.py b/comfy/text_encoders/t5.py
index ba628478d568..a1420c6cd2f2 100644
--- a/comfy/text_encoders/t5.py
+++ b/comfy/text_encoders/t5.py
@@ -1,6 +1,7 @@
 import torch
 import math
 from comfy.ldm.modules.attention import optimized_attention_for_device
+import comfy.ops
 
 class T5LayerNorm(torch.nn.Module):
     def __init__(self, hidden_size, eps=1e-6, dtype=None, device=None, operations=None):
@@ -11,7 +12,7 @@ def __init__(self, hidden_size, eps=1e-6, dtype=None, device=None, operations=No
     def forward(self, x):
         variance = x.pow(2).mean(-1, keepdim=True)
         x = x * torch.rsqrt(variance + self.variance_epsilon)
-        return self.weight.to(device=x.device, dtype=x.dtype) * x
+        return comfy.ops.cast_to_input(self.weight, x) * x
 
 activations = {
     "gelu_pytorch_tanh": lambda a: torch.nn.functional.gelu(a, approximate="tanh"),
@@ -82,7 +83,7 @@ def __init__(self, model_dim, inner_dim, num_heads, relative_attention_bias, dty
         if relative_attention_bias:
             self.relative_attention_num_buckets = 32
             self.relative_attention_max_distance = 128
-            self.relative_attention_bias = torch.nn.Embedding(self.relative_attention_num_buckets, self.num_heads, device=device)
+            self.relative_attention_bias = operations.Embedding(self.relative_attention_num_buckets, self.num_heads, device=device, dtype=dtype)
 
     @staticmethod
     def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128):
@@ -132,7 +133,7 @@ def _relative_position_bucket(relative_position, bidirectional=True, num_buckets
         relative_buckets += torch.where(is_small, relative_position, relative_position_if_large)
         return relative_buckets
 
-    def compute_bias(self, query_length, key_length, device):
+    def compute_bias(self, query_length, key_length, device, dtype):
         """Compute binned relative position bias"""
         context_position = torch.arange(query_length, dtype=torch.long, device=device)[:, None]
         memory_position = torch.arange(key_length, dtype=torch.long, device=device)[None, :]
@@ -143,7 +144,7 @@ def compute_bias(self, query_length, key_length, device):
             num_buckets=self.relative_attention_num_buckets,
             max_distance=self.relative_attention_max_distance,
         )
-        values = self.relative_attention_bias(relative_position_bucket)  # shape (query_length, key_length, num_heads)
+        values = self.relative_attention_bias(relative_position_bucket, out_dtype=dtype)  # shape (query_length, key_length, num_heads)
         values = values.permute([2, 0, 1]).unsqueeze(0)  # shape (1, num_heads, query_length, key_length)
         return values
 
@@ -152,7 +153,7 @@ def forward(self, x, mask=None, past_bias=None, optimized_attention=None):
         k = self.k(x)
         v = self.v(x)
         if self.relative_attention_bias is not None:
-            past_bias = self.compute_bias(x.shape[1], x.shape[1], x.device)
+            past_bias = self.compute_bias(x.shape[1], x.shape[1], x.device, x.dtype)
 
         if past_bias is not None:
             if mask is not None:
@@ -199,7 +200,7 @@ def __init__(self, num_layers, model_dim, inner_dim, ff_dim, ff_activation, gate
         self.final_layer_norm = T5LayerNorm(model_dim, dtype=dtype, device=device, operations=operations)
         # self.dropout = nn.Dropout(config.dropout_rate)
 
-    def forward(self, x, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True):
+    def forward(self, x, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
@@ -225,7 +226,7 @@ def __init__(self, config_dict, dtype, device, operations):
 
         self.encoder = T5Stack(self.num_layers, model_dim, model_dim, config_dict["d_ff"], config_dict["dense_act_fn"], config_dict["is_gated_act"], config_dict["num_heads"], config_dict["model_type"] != "umt5", dtype, device, operations)
         self.dtype = dtype
-        self.shared = torch.nn.Embedding(config_dict["vocab_size"], model_dim, device=device)
+        self.shared = operations.Embedding(config_dict["vocab_size"], model_dim, device=device, dtype=dtype)
 
     def get_input_embeddings(self):
         return self.shared
@@ -234,5 +235,7 @@ def set_input_embeddings(self, embeddings):
         self.shared = embeddings
 
     def forward(self, input_ids, *args, **kwargs):
-        x = self.shared(input_ids)
+        x = self.shared(input_ids, out_dtype=kwargs.get("dtype", torch.float32))
+        if self.dtype not in [torch.float32, torch.float16, torch.bfloat16]:
+            x = torch.nan_to_num(x) #Fix for fp8 T5 base
         return self.encoder(x, *args, **kwargs)
diff --git a/comfy/utils.py b/comfy/utils.py
index 1e4b5ef882d2..0db9fbb62673 100644
--- a/comfy/utils.py
+++ b/comfy/utils.py
@@ -11,7 +11,7 @@
 def load_torch_file(ckpt, safe_load=False, device=None):
     if device is None:
         device = torch.device("cpu")
-    if ckpt.lower().endswith(".safetensors"):
+    if ckpt.lower().endswith(".safetensors") or ckpt.lower().endswith(".sft"):
         sd = safetensors.torch.load_file(ckpt, device=device.type)
     else:
         if safe_load:
diff --git a/folder_paths.py b/folder_paths.py
index 2baf8ce1c9f2..71faa2df4db7 100644
--- a/folder_paths.py
+++ b/folder_paths.py
@@ -3,7 +3,7 @@
 import logging
 from typing import Set, List, Dict, Tuple
 
-supported_pt_extensions: Set[str] = set(['.ckpt', '.pt', '.bin', '.pth', '.safetensors', '.pkl'])
+supported_pt_extensions: Set[str] = set(['.ckpt', '.pt', '.bin', '.pth', '.safetensors', '.pkl', '.sft'])
 
 SupportedFileExtensionsType = Set[str]
 ScanPathType = List[str]
diff --git a/nodes.py b/nodes.py
index dc66dd3ef00c..5b6321b69d72 100644
--- a/nodes.py
+++ b/nodes.py
@@ -818,15 +818,17 @@ class UNETLoader:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "unet_name": (folder_paths.get_filename_list("unet"), ),
+                              "weight_dtype": (["default", "fp8_e4m3fn", "fp8_e5m2"],)
                              }}
     RETURN_TYPES = ("MODEL",)
     FUNCTION = "load_unet"
 
     CATEGORY = "advanced/loaders"
 
-    def load_unet(self, unet_name):
+    def load_unet(self, unet_name, weight_dtype):
+        weight_dtype = {"default":None, "fp8_e4m3fn":torch.float8_e4m3fn, "fp8_e5m2":torch.float8_e4m3fn}[weight_dtype]
         unet_path = folder_paths.get_full_path("unet", unet_name)
-        model = comfy.sd.load_unet(unet_path)
+        model = comfy.sd.load_unet(unet_path, dtype=weight_dtype)
         return (model,)
 
 class CLIPLoader:
@@ -859,7 +861,7 @@ class DualCLIPLoader:
     def INPUT_TYPES(s):
         return {"required": { "clip_name1": (folder_paths.get_filename_list("clip"), ),
                               "clip_name2": (folder_paths.get_filename_list("clip"), ),
-                              "type": (["sdxl", "sd3"], ),
+                              "type": (["sdxl", "sd3", "flux"], ),
                              }}
     RETURN_TYPES = ("CLIP",)
     FUNCTION = "load_clip"
@@ -873,6 +875,8 @@ def load_clip(self, clip_name1, clip_name2, type):
             clip_type = comfy.sd.CLIPType.STABLE_DIFFUSION
         elif type == "sd3":
             clip_type = comfy.sd.CLIPType.SD3
+        elif type == "flux":
+            clip_type = comfy.sd.CLIPType.FLUX
 
         clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type)
         return (clip,)
@@ -1843,6 +1847,7 @@ def expand_image(self, image, left, top, right, bottom, feathering):
     "StyleModelLoader": "Load Style Model",
     "CLIPVisionLoader": "Load CLIP Vision",
     "UpscaleModelLoader": "Load Upscale Model",
+    "UNETLoader": "Load Diffusion Model",
     # Conditioning
     "CLIPVisionEncode": "CLIP Vision Encode",
     "StyleModelApply": "Apply Style Model",