UPDATE V1.3: Not a revolution but tries to find a middle ground between v1 and v1.2. I hope it fixes the biggest issues people had with both versions - we shall see. Both base (bf16) and fp8 have been uploaded (fp8 on the right ----->).
Chroma is a fantastic and highly versatile model capable of producing photo-like results, but it requires careful prompting. This finetune aims to improve reliability in realistic/photo-based styles while preserving Chroma’s broad concept knowledge. The v1.3 flash version has the rank-256 lora (from here) baked in (the other flash versions use rank-128). GGUFs on HuggingFace.
Prompting: Chroma prompts work well. Simple prompts describing what you want to see in natural sentences works well. Some example images show the captioning style used in training. Negative prompts do not work with CFG set to one. With CFG above one, negative prompts work and can be very important (for good or bad).
Example settings (not necessarily optimal):
Workflow: Chroma template workflow in ComfyUI
Steps (base): ~30-40 (depends on other settings; CFG, sampler, etc.)
Steps (flash lora): 15-17 works well with rank-128/256. Depends on lora rank.
CFG (base): ~3.5 (depends on other settings; steps, sampler, etc.)
CFG (flash lora): 1 works well with rank-128/256. Depends on lora rank.
Sampler: res_2m and dpmpp_sde work well.
Scheduler: I like bong_tangent | beta is also good.
Note on settings: If you change one setting (sampler, CFG, steps) you probably have to change others to get good results. CFG affects speed.
Support: Have too much money? Want to support further training? https://ko-fi.com/dawncreates
Training Details The model was trained locally, using Chroma-HD as the base. Each epoch included images at 3–5 different resolutions, though only a subset of the dataset was used per epoch. Except for the extra resolutions, OneTrainer's default config for 24gb Chroma finetuning was used. The dataset consists almost exclusively of SFW-images of people and landscapes, so to retain Chroma-HD's original conceptual understanding, several layers were merged back at various ratios. All the juice, compositions, subjects, and concepts come from Chroma itself, my model just nudges it towards realism. Honestly, this version is more of a showcase of how good Chroma is than a great finetune in itself. I do think it shows how much potential Chroma has for finetuning though - so get to work on Chroma finetuners - it has so much potential!
All images were captioned using JoyCaption: https://github.com/fpgaminer/joycaption
The model was trained using OneTrainer: https://github.com/Nerogar/OneTrainer