I've made this model using the <a target="_blank" rel="ugc" href="https://github.com/lum3on/ComfyUI-ModelQuantizer">https://github.com/lum3on/ComfyUI-ModelQuantizer</a> nodes and the full V48 version, so it could run better in my RX6800.I know that it is also needed if you want to run TorchCompile in RTX 3000 series, so here you have it.

Chroma fp8 e5m2

Quantizer from the model provided by Phr00t in huggingface - <a target="_blank" rel="ugc" href="https://huggingface.co/Phr00t/Chroma-Rapid-AIO">https://huggingface.co/Phr00t/Chroma-Rapid-AIO</a>Use a CFG of 1, 8 steps and sa_solver sampler with the beta scheduler (V3). However, the "ipndm" sampler gives good results too, particularly more realistic ones.

Files with SHA256: 43e7e6b74357c6dc45cc03943bbae624cda4afcf6429ed23e30e5f90f4cf8723