diff --git a/README.md b/README.md index ede46580..69b84a15 100644 --- a/README.md +++ b/README.md @@ -7,20 +7,19 @@ I am going to keep updating it with new features. Suggestions are welcome. To use: git clone https://github.com/maybleMyers/chromaforge -Download one of the t5xxl text encoders from https://huggingface.co/silveroxides/t5xxl_flan_enc and place in models/text_encoder. +Download one of the t5xxl text encoder from https://huggingface.co/flux-safetensors/flux-safetensors/blob/main/t5xxl_fp16.safetensors and place in models/text_encoder. Download the vae from https://huggingface.co/lodestones/Chroma/tree/main ae.safetenstors and put it in models/vae. -Download a checkpoint from https://huggingface.co/lodestones/Chroma/tree/main ie chroma-unlocked-v29.5.safetensors and place in models/Stable-diffusion. +Download the main chroma model from https://huggingface.co/lodestones/Chroma1-HD/blob/main/Chroma1-HD.safetensors You will need python3.10 installed. Run webui-user.bat for windows installation. Run webui.sh for linux installation (./webui.sh) from the chromaforge root directory. After you first run webui-user.bat it might error, if so, close it and re run it. To update to the latest version navigate to your root directory in a terminal and type "git pull" -in Forge, on the top left select all and not flux +in Forge, on the top left select chroma select the checkpoint and then in the next field select the text encoder and vae. -Use euler simple scheduler. -set the distilled config scale to 1, and the normal config scale to something like 3.5-7 -use a very long positive prompt and a very long negative prompt. +Use euler simple scheduler to start. +set the CFG scale to something like 3.5-7 forge doesn't seem to work with all quantized model, Q4_K_S fail, but Q4_1 work refer to screenshot for working settings. @@ -34,28 +33,23 @@ there's a bunch of extra samplers/schedulers at these places: https://github.com/DenOfEquity/webUI_ExtraSchedulers https://github.com/MisterChief95/sd-forge-extra-samplers -Here are experimental high resolution checkpoints for chroma: -https://huggingface.co/lodestones/chroma-debug-development-only/tree/main/staging_large_3 -https://huggingface.co/lodestones/chroma-debug-development-only/tree/main/staging_large_4 -They can be converted to safetensors with convertpth.py -for example modify the source.pth and output.safetensors on the last line in the script: convert_pth_to_safetensors("source.pth", "output.safetensors") to what you want them to be, then run the script with python convertpth.py. - ## RamForge -This supports use of the RamTorch memory management system developed by lord Loadstones. Add --use-ramtorch-chroma to your command line arguments. It will drastically lower vram usage to let you increase batch size or resolution greatly above what you could previously use. For instance you can run a batch of 8 1024x1024 images with a 12gb GPU. It does increas normal RAM usage by alot, 64GB recommended. +This supports use of the RamTorch memory management system developed by lord Loadstones. Add --use-ramtorch-chroma to your command line arguments. It will drastically lower vram usage to let you increase batch size or resolution greatly above what you could previously use. For instance you can run a batch of 8 1024x1024 images with a 12gb GPU. It does increas normal RAM usage by alot, 64GB recommended. https://github.com/lodestone-rock/RamTorch ## Radiance Model Download a checkpoint from here https://huggingface.co/lodestones/chroma-debug-development-only/tree/main/radiance then convert it with convertpth.py. Put it in your stable diffusion subfolder. Use the radiance model in the chroma tab of the gui without a vae. It will be automatically detected. Use the same T5 you normally use. -The memory management is currently derped. It will run fast with a 24GB card when setting the gpu weights to about 19,200, but if you need to offload expect 4x slowdown. +The memory management is currently derped. It will run fast with a 24GB card when setting the gpu weights to about 19,200, but if you need to offload expect 4x slowdown. You can dick with the gpu weights setting and it might run good if you set it to like half your vram. Training: https://github.com/lodestone-rock/flow/ https://github.com/kohya-ss/sd-scripts/tree/sd3 Donate to Lodestone (training is bookoo expensive and crowdfunded): https://ko-fi.com/lodestonerock Discord: http://discord.gg/SQVcWVbqKx ## Changlog + 9/19/2025 Added support for RamTorch memory management system using normal chroma models. Use it by moidfying your webui-user.bat's set COMMANDLINE_ARGS= --use-ramtorch-chroma https://github.com/lodestone-rock/RamTorch 9/2/2025