Add flux-image-variations

flux-image-variations/config.json

@@ -0,0 +1,19 @@
+{
+  "_class_name": "FluxTransformer2DModel",
+  "_diffusers_version": "0.31.0",
+  "_name_or_path": "output-model",
+  "attention_head_dim": 128,
+  "axes_dims_rope": [
+    16,
+    56,
+    56
+  ],
+  "guidance_embeds": false,
+  "in_channels": 64,
+  "joint_attention_dim": 4096,
+  "num_attention_heads": 24,
+  "num_layers": 19,
+  "num_single_layers": 38,
+  "patch_size": 1,
+  "pooled_projection_dim": 768
+}

flux-image-variations/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6bfb9d8e1fdfa88ef5b6290d4bbb4c4b6c5db1c1d16d35f33153230b6b0554e
+size 47564852120

inference_flux_model.py → inference_flux.py

@@ -71,7 +71,7 @@ def main():
     global pipe
     clip = CLIPModel.from_pretrained("openai/clip-vit-large-patch14", torch_dtype=torch.bfloat16)
     pipe = FluxWithCFGPipeline.from_pretrained("ostris/OpenFLUX.1", text_encoder=clip, transformer=None, torch_dtype=torch.bfloat16)
-    pipe.transformer = FluxTransformer2DModel.from_pretrained("flux-image-variations-model", torch_dtype=torch.bfloat16)
+    pipe.transformer = FluxTransformer2DModel.from_pretrained("flux-image-variations", torch_dtype=torch.bfloat16)
     pipe.to("cuda")
 
     img_prompt = Image.open(args.image_prompt) if args.image_prompt else None
@@ -79,4 +79,4 @@ def main():
              num_images=args.num_images, resolution=args.resolution)
 
 if __name__ == "__main__":
-    main()
+    main()

open_flux_pipeline.py

@@ -0,0 +1,242 @@
+import numpy as np
+import torch
+from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput
+from diffusers import FluxPipeline
+from typing import List, Union, Optional, Dict, Any, Callable
+from diffusers.pipelines.flux.pipeline_flux import calculate_shift, retrieve_timesteps
+
+from diffusers.utils import is_torch_xla_available
+
+from modified_flux import FluxImageConditionedPipeline
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+# TODO this is rough. Need to properly stack unconditional or make it optional
+class FluxWithCFGPipeline(FluxImageConditionedPipeline):
+    def __call__(
+        self,
+        image_prompt = None,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        negative_image_prompt = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 28,
+        timesteps: List[int] = None,
+        guidance_scale: float = 7.0,
+        num_images_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+    ):
+
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            height,
+            width,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            max_sequence_length=max_sequence_length,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._joint_attention_kwargs = joint_attention_kwargs
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        lora_scale = (
+            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
+        )
+        (
+            prompt_embeds,
+            pooled_prompt_embeds,
+            text_ids,
+        ) = self.encode_prompt(
+            image_prompt=image_prompt,
+            prompt=prompt,
+            prompt_2=prompt_2,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+            lora_scale=lora_scale,
+        )
+        (
+            negative_prompt_embeds,
+            negative_pooled_prompt_embeds,
+            negative_text_ids,
+        ) = self.encode_prompt(
+            image_prompt=negative_image_prompt,
+            prompt=negative_prompt,
+            prompt_2=negative_prompt_2,
+            prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+            lora_scale=lora_scale,
+        )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels // 4
+        latents, latent_image_ids = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 4.5. Concat CLIP image embeddings
+        # if image_prompt is not None:
+        #     image_prompt_embeds = self._encode_image_with_clip(image_prompt, device=device)
+        #     latents_conditioned = torch.cat((latents, image_prompt_embeds.reshape((batch_size, 12, 64))), dim=1)
+        #     latent_image_ids_conditioned = torch.cat((latent_image_ids, torch.zeros(12, 3).to(device)), dim=0)
+        # else:
+        #     latents_conditioned = latents
+        #     latent_image_ids_conditioned = latent_image_ids
+
+        # 5. Prepare timesteps
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+        image_seq_len = latents.shape[1]
+        mu = calculate_shift(
+            image_seq_len,
+            self.scheduler.config.base_image_seq_len,
+            self.scheduler.config.max_image_seq_len,
+            self.scheduler.config.base_shift,
+            self.scheduler.config.max_shift,
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas,
+            mu=mu,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+                print(f" - Memory: {torch.cuda.memory_allocated()/(1024**3):.3f} GB")
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+                # handle guidance
+                if self.transformer.config.guidance_embeds:
+                    guidance = torch.tensor([guidance_scale], device=device)
+                    guidance = guidance.expand(latents.shape[0])
+                else:
+                    guidance = None
+
+                noise_pred_text = self.transformer(
+                    hidden_states=latents,#_conditioned,
+                    timestep=timestep / 1000,
+                    guidance=guidance,
+                    pooled_projections=pooled_prompt_embeds,
+                    encoder_hidden_states=prompt_embeds,
+                    txt_ids=text_ids,
+                    img_ids=latent_image_ids,#_conditioned,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )[0]
+                # if image_prompt is not None:
+                #     noise_pred_text = noise_pred_text.narrow(1, 0, 1024)
+
+                # todo combine these
+                noise_pred_uncond = self.transformer(
+                    hidden_states=latents,
+                    timestep=timestep / 1000,
+                    guidance=guidance,
+                    pooled_projections=negative_pooled_prompt_embeds,
+                    encoder_hidden_states=negative_prompt_embeds,
+                    txt_ids=negative_text_ids,
+                    img_ids=latent_image_ids,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                print(f" - Memory: {torch.cuda.memory_allocated()/(1024**3):.3f} GB")
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if output_type == "latent":
+            image = latents
+
+        else:
+            latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
+            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+            image = self.vae.decode(latents.to(device, dtype=self.vae.dtype), return_dict=False)[0]
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return FluxPipelineOutput(images=image)