Execute the same positive prompt, negative prompt, steps, denoise and cfg with every possible combination of samplers and schedulers. 306 KSamplers in total.
All images are saved into 306 different folders with this path format: output/hidream/sampler_scheduler/sampler_scheduler_number_.png
You can use the python script to generate composite images like the ones shown as examples. Copy the combine.py into the output/hidream folder. Edit the python script and change the line "image_gen_number = "00001" to whatever the number of your generation is. Youll have to use pip to install some packages but i forgot which ones.
Youll have to reopen the workflow after every use because comfyui changes all "int" nodes above 2048 to 2048. so you can only execute that workflow one time basically but its a not an issue since running the entire workflow takes 10 hours on a 5090.
I recommend deactivating the hardware acceleration in your browser while executing this workflow or itll use a lot of vram.
Also yes i realize theres better and easier ways of doing the exact same thing.
Description
FAQ
Comments (7)
Thanks for sharing this! I’d really like to know what you think is the best mix for getting a realistic approach.
i dont think theres one combination thats best for realism in general but these tests show that some combinations dont work at all. Going with whatever sampler is fastest and trying different schedulers with it is probably a good approach because it allows you to generate more images which gives you a better chance of getting a good image.
Thanks a lot for these tests.
What version of HiDream are these tests targeting? FULL、DEV or FAST？
Models ive used for these tests:
https://huggingface.co/Comfy-Org/HiDream-I1_ComfyUI/blob/main/split_files/diffusion_models/hidream_i1_full_fp16.safetensors
https://huggingface.co/Comfy-Org/HiDream-I1_ComfyUI/blob/main/split_files/text_encoders/clip_l_hidream.safetensors
https://huggingface.co/Comfy-Org/HiDream-I1_ComfyUI/blob/main/split_files/text_encoders/clip_g_hidream.safetensors
https://huggingface.co/easygoing0114/flan-t5-xxl-fused/blob/main/flan_t5_xxl_full_fp32.safetensors
https://huggingface.co/aifeifei798/DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored/tree/main (combined)
They were ran on a 5090 and 192gb of cpu ram. Each run took ~10 hours
Thanks for those, would have appreciated the two prompts to get the alignment, and as Metafilm said, which Hidream. Anyway, thanks a lot for this, and your time.
i really liked hidream's prompt adherence, so i started to play around with it. And built a tool to compare the generations along the way.
https://aicompare-85aun.sevalla.page/
Its in progress, have a couple of more ideas to implement. for now, it recognizes the base hidream workflow.