The grid artifacts are fixed! I'm stunned that it's so easy to get a model to generate such high quality pictures, from scratch, in less than a week, on mid range hardware from 10 years ago. AI is truly amazing. I thought it would be like pulling teeth, but no, it went very smoothly. Just asked Deepseek to build it layer by layer, kernel by kernel and it just worked.
user@desktop:~/image-generation-rx580/gcn-code$ time ./build/run-h3 ../weights/MiniMax-H3-encoder-Q4_K_M.gguf ../weights/MiniMax-H3-Image-AD-Q5_K.gguf ../weights/minimax_h3_image_vae_fp32.safetensors "A silver tabby kitten holding a sign reading '/vcg/', natural daylight" ~/Pictures/vcg-cat.png 1024
Processing prompt.
Layers:
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49
dit: 50 layers, RAM-retained [0,19) 4076 MiB, VRAM-resident 16 blocks 3945 MiB, streamed 15
step 0: sigma 1.0000 -> 0.9675 latent rms 1.0014 v rms 1.2772
step 1: sigma 0.9675 -> 0.8875 latent rms 0.9689 v rms 1.3377
step 2: sigma 0.8875 -> 0.7500 latent rms 0.8952 v rms 1.3283
step 3: sigma 0.7500 -> 0.5328 latent rms 0.7894 v rms 1.2863
step 4: sigma 0.5328 -> 0.2320 latent rms 0.6971 v rms 1.2425
step 5: sigma 0.2320 -> 0.0000 latent rms 0.7658 v rms 1.0970
wrote /home/user/Pictures/vcg-cat.png (1024x1024)
real 5m18,684s
user 0m11,004s
sys 0m33,353s
user@desktop:~/image-generation-rx580/gcn-code$ find . -type f -name '*.c' -exec cat {} + | wc -l
11867
user@desktop:~/image-generation-rx580/gcn-code$ find . -type f -name '*.cl' -exec cat {} + | wc -l
3598