black/ blur image when use FLUX-dev-INT4 with

[ak-sangal]

i am using tayler_text2image.py script with model FLUX-dev-INT4 but it generate black/ distorted image . this script generates both base line image- Black and taylerseer script-based image blurry. i have intel irisx igpu with 16 GB RAM i71355U processor with window 11, below is logs

(ov_env) E:\Openvino _Ai>python taylorseer.py .\Model\FLUX-dev-INT4 "A woman stands alone in an overgrown grass field on a late night, surrounded by darkness that only adds to her allure. Her long straight brunette hair cascades down her back, framing her face with elegance. A light-green t-shirt clings to her body, showcasing her toned physique and sensuality. Denim shorts hang loose around her midriff, giving off an air of comfort rather than style. Her eyes sparkle with laughter, as a bright smile spreads across her face. The bokeh effects blur the background, adding an intimate and contemplative feel to the moment. The dark surroundings seem to envelop her, creating a sense of mystery and adventure that's almost palpable."

Generating baseline image without caching...
Step 1/28
Step 2/28
Step 3/28
Step 4/28
Step 5/28
Step 6/28
Step 7/28
Step 8/28
Step 9/28
Step 10/28
Step 11/28
Step 12/28
Step 13/28
Step 14/28
Step 15/28
Step 16/28
Step 17/28
Step 18/28
Step 19/28
Step 20/28
Step 21/28
Step 22/28
Step 23/28
Step 24/28
Step 25/28
Step 26/28
Step 27/28
Step 28/28
Baseline generation completed in 315.85s
Baseline image saved to taylorseer_baseline.bmp

Generating image with TaylorSeer caching...
TaylorSeerCacheConfig {
cache_interval: 3
disable_cache_before_step: 6
disable_cache_after_step: -2
}
Step 1/28
Step 2/28
Step 3/28
Step 4/28
Step 5/28
Step 6/28
Step 7/28
Step 8/28
Step 9/28
Step 10/28
Step 11/28
Step 12/28
Step 13/28
Step 14/28
Step 15/28
Step 16/28
Step 17/28
Step 18/28
Step 19/28
Step 20/28
Step 21/28
Step 22/28
Step 23/28
Step 24/28
Step 25/28
Step 26/28
Step 27/28
Step 28/28
TaylorSeer generation completed in 161.06s
Image saved to taylorseer.bmp

Performance Comparison:
Baseline time: 315.85s
TaylorSeer time: 161.06s
Speedup: 1.96x
Time saved: 154.79s (49.0%)

[l-bat]

Hi @ak-sangal!

Could you please verify if you can get a correct image using the original FP FLUX-dev model on iGPU?
Also, could you check if the FLUX-dev-INT4 model produces correct results when running on CPU?

[ak-sangal]

When I use cpu it works best but I want to use gpu on cpu it's very slow

[ak-sangal]

On same pc when I run OpenVINO/FLUX.1-schnell-int4-ov it's works both int4 and int8 model

[l-bat]

@as-suvorov please reassign to GPU team. Looks like that the issue is not related to the TaylorSeer algorithm.

[geunhwan]

@ak-sangal Just a question. In your IR xml, can you find the following value?

<ACTIVATIONS_SCALE_FACTOR value="8.0" />

[ak-sangal]

D:\Openvino model\openvino model\FLUX-dev-INT4\text_encoder_2\openvino_model.xml: <ACTIVATIONS_SCALE_FACTOR value="8.0" />
D:\Openvino model\openvino model\FLUX-dev-INT4\transformer\openvino_model.xml: <ACTIVATIONS_SCALE_FACTOR value="8.0" />
D:\Openvino model\openvino model\FLUX-dev-INT4\vae_decoder\openvino_model.xml: <ACTIVATIONS_SCALE_FACTOR value="8.0" />
D:\Openvino model\openvino model\FLUX-dev-INT4\vae_encoder\openvino_model.xml: <ACTIVATIONS_SCALE_FACTOR value="8.0" />