refiners/tests/e2e/test_diffusion_ref/README.md
2023-09-13 14:24:53 +02:00

4.8 KiB

Note about this data

Expected outputs

expected_*.png files are the output of the same diffusion run with a different codebase, usually diffusers with the same settings as us (DPMSolverMultistepScheduler, VAE patched to remove randomness, same seed...).

For instance here is how we generate expected_std_random_init.png:

import torch

from diffusers import DPMSolverMultistepScheduler
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    torch_dtype=torch.float32,
).to("cuda")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)

prompt = "a cute cat, detailed high-quality professional image"
negative_prompt = "lowres, bad anatomy, bad hands, cropped, worst quality"

torch.manual_seed(2)
output = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=30,
    guidance_scale=7.5,
)

output.images[0].save("std_random_init_expected.png")

Special cases:

  • expected_refonly.png has been generated with Stable Diffusion web UI.
  • expected_inpainting_refonly.png, expected_image_ip_adapter_woman.png, expected_image_sdxl_ip_adapter_woman.png and expected_ip_adapter_controlnet.png have been generated with refiners itself (and inspected so that they look reasonable).

Other images

VAE without randomness

--- a/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py
+++ b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py
@@ -524,13 +524,8 @@ class StableDiffusionImg2ImgPipeline(DiffusionPipeline):
                 f" size of {batch_size}. Make sure the batch size matches the length of the generators."
             )

-        if isinstance(generator, list):
-            init_latents = [
-                self.vae.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size)
-            ]
-            init_latents = torch.cat(init_latents, dim=0)
-        else:
-            init_latents = self.vae.encode(image).latent_dist.sample(generator)
+        init_latents = [self.vae.encode(image[i : i + 1]).latent_dist.mean for i in range(batch_size)]
+        init_latents = torch.cat(init_latents, dim=0)

         init_latents = self.vae.config.scaling_factor * init_latents

Textual Inversion

  • expected_textual_inversion_random_init.png has been generated with StableDiffusionPipeline, e.g.:
import torch

from diffusers import DPMSolverMultistepScheduler
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    torch_dtype=torch.float32,
).to("cuda")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
pipe.load_textual_inversion("sd-concepts-library/gta5-artwork")

prompt = "a cute cat on a <gta5-artwork>"
negative_prompt = ""

torch.manual_seed(2)
output = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=30,
    guidance_scale=7.5,
)

output.images[0].save("expected_textual_inversion_random_init.png")