Test result is not pixel-level consistency

[blacklong28]

Hi, I have test your fast easycontrol, and vs origin easycontrol, is not maintaining pixel-level consistency.
I copy and replace your code with https://github.com/jax-explorer/ComfyUI-easycontrol.
There has indeed been some acceleration, but is not maintaining pixel-level consistency.
Why you say maintaining pixel-level consistency, and how did you test for consistency?
origin：

fast：

[blacklong28]

Use your infer code, origin vs fast:

[KBRASK]

Hello! My code is pixel-level identical to the one at https://github.com/Xiaojiu-z/EasyControl. I'm not sure what modifications have been made in the ComfyUI version. Additionally, please ensure that both versions are run on the same machine, with the same environment and random seed. If no random seed is set, it will be chosen randomly, which means that even the same algorithm can produce different results. Please be aware of this.
In addition, I conducted extensive comparisons, both step-by-step and at the logical level. The generation results and underlying logic in my implementation are consistent with those of EasyControl. As a result, I achieved pixel-level identical outputs and conclusions.

[blacklong28]

@KBRASK Hello, I test with your infer code:

def test_fast():
    base_path = "black-forest-labs/FLUX.1-dev"

    pipe = FluxPipeline_fast.from_pretrained(base_path, torch_dtype=torch.bfloat16).to("cuda")

    set_single_lora(pipe.transformer, "Ghibli.safetensors", lora_weights=[1], cond_size=512)

    prompt = "A woman"

    url = "https://github.com/ToTheBeginning/PuLID/blob/main/example_inputs/liuyifei.png?raw=true"
    response = requests.get(url)
    spatial_image = Image.open(BytesIO(response.content))
    width, height = spatial_image.size
    new_width = (width // 8) * 8
    new_height = (height // 8) * 8

    image = pipe(
        prompt,
        height=new_height,
        width=new_width,
        guidance_scale=3.5,
        num_inference_steps=25,
        max_sequence_length=512,
        generator=torch.Generator("cpu").manual_seed(5),
        spatial_images=[spatial_image],
        subject_images=[],
        cond_size=512,
    ).images[0]

    # Clear cache after generation
    clear_cache(pipe.transformer)

    image.save("output2_fast.png")

and test origin easycontrol by follow code:

import torch
from PIL import Image
from src.pipeline import FluxPipeline
from src.transformer_flux import FluxTransformer2DModel
from src.lora_helper import set_single_lora, set_multi_lora
import requests
from io import BytesIO

def clear_cache(transformer):
    for name, attn_processor in transformer.attn_processors.items():
        attn_processor.bank_kv.clear()
# Initialize model
device = "cuda"
base_path = "black-forest-labs/FLUX.1-dev"  # Path to your base model
pipe = FluxPipeline.from_pretrained(base_path, torch_dtype=torch.bfloat16, device=device)
transformer = FluxTransformer2DModel.from_pretrained(
    base_path, 
    subfolder="transformer",
    torch_dtype=torch.bfloat16, 
    device=device
)
pipe.transformer = transformer
pipe.to(device)

set_single_lora(pipe.transformer, "Ghibli.safetensors", lora_weights=[1], cond_size=512)
prompt = "A woman"

url = "https://github.com/ToTheBeginning/PuLID/blob/main/example_inputs/liuyifei.png?raw=true"
response = requests.get(url)
spatial_image = Image.open(BytesIO(response.content))
width, height = spatial_image.size
new_width = (width // 8) * 8
new_height = (height // 8) * 8

image = pipe(
    prompt,
    height=new_height,
    width=new_width,
    guidance_scale=3.5,
    num_inference_steps=25,
    max_sequence_length=512,
    generator=torch.Generator("cpu").manual_seed(5),
    spatial_images=[spatial_image],
    subject_images=[],
    cond_size=512,
).images[0]

# Clear cache after generation
clear_cache(pipe.transformer)

image.save("output_origin.png")

The results are still not consistent.

I already use the origin code from https://github.com/Xiaojiu-z/EasyControl, and the ComfyUI version vs the https://github.com/Xiaojiu-z/EasyControl version now I test is pixel-level consistency.

In addition, I have found that in your code

FastEasyControl/src/layers_cache.py

Lines 360 to 363 in 37ae093

	# Linear projection (with LoRA weight applied to each proj layer)
	hidden_states = attn.to_out[0](hidden_states)
	encoder_hidden_states = attn.to_add_out(encoder_hidden_states)

the calculation of proj_loras is missing compared to the original implementation code.
The fast fix in the above picture is the result after I added it, and there are still some inconsistencies. I would like to know how you compare the results and whether it is on the latest code. Could you provide the test code?
Thanks!