Skip to content

Conversation

liutyi
Copy link
Contributor

@liutyi liutyi commented Aug 26, 2025

Description

Some more model covers.

Notes

Nothing critical. Maybe merged right now. or tomorrow or couple of days later (with some more covers)

Environment and Testing

local

Fun facts (if any)

  1. Qwen Image Edit 1024x1024@50STEP txt2img takes about 4h (bf16, AG4, ipex, arc iGPU)
  2. Switching between models like Qwen and aMUSEd is kind of impressive. like what path generative Ai done in couple of years.
  3. Alternatives that was generated but not selected as covers and prompts are there

@vladmandic
Copy link
Owner

lol!
for the slowest ones, flag them and send me exact prompts and i can run them on my system.

@liutyi
Copy link
Contributor Author

liutyi commented Aug 26, 2025

lol! for the slowest ones, flag them and send me exact prompts and i can run them on my system.

unfortunately, it is not work like that. I actually have no idea which prompt will be final for each model. I try to keep some resemblance with original covers (if it possible). And also want somehow showcase the model features. And size.. And group model families at least by background colors. It is kind multi iteration trying process. Sometimes starting with 2-10 test images to find out better CFG/STEPS or working resolution. So If I knew exact prompts, I would just set scheduler to process all that at night.

It is actually one Model/Prompt I would like you to run:

it is Lumina-Image-2.0 1024x1024 Prompt:
A detail photo-realistic image of a steampunk locomotive on a platform with sharp lines, surrounded by light purple fog

because what I got there is just wrong

2025-08-26_220602

@vladmandic
Copy link
Owner

lumina-2.0 with your prompt...
00006-Lumina-Image-2 0

@vladmandic vladmandic merged commit 4af30f5 into vladmandic:dev Aug 27, 2025
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants