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. bf16 and fp8 available --->. GGUFs on HuggingFace.
Prompting: Simple prompts describing what you want to see in natural language works well (example1, example2). Chroma prompts work well. Examples of captioning style used in training: woman sitting, waterfall, wolf, woman in dessert. Negative prompts don't work at CFG one, but above one negative prompts can be important.
Example settings (not necessarily optimal) using Comfy's default Chroma workflow:
Base: Steps:: ~30-40. CFG: ~3.5 (best settings depend on steps/CFG/sampler, etc.).
Flash (lora rank-128 or 256): Steps: 15-17. CFG 1. (Depends on lora rank, sampler, etc.)
Sampler: Examples use res_2m, dpmpp_sde, or exp_heun_2_x0_sde. Others are also good.
Scheduler: I like bong_tangent | beta & beta57 and others are also good
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 a showcase of how good Chroma is. 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
Description
FAQ
Comments (20)
Love general realism improvement Lora's.
Need to check this out.
The low-step one is not Apache-2.0 Licensed. By merging that LoRA in to it which has Flux1-dev trained weights you effectively forced it to become Licensed under same license as it.
FLUX.1 [dev] Non-Commercial License
Isn't it a Chroma lora? I thought it was. I'll take down the lowstep model for now and check / test other loras. Thanks for the heads up.
Your flash loras are apache, correct? Anyway, I'll do some tests with them when I have time and link to them with recommended settings. Thanks again.
@dawncreates Don't care, he deleted some loras because they weren't licensed under Apache 2 by their original creators
@dawncreates Yeah it was made by pruning layers of a LoRA extracted from a Flux1 Hyper model similar to ByteDance/HyperSD. Yeah you can use pretty much any LoRA with flash in name plus the one named few-steps without issue. It would not be an issue if you used any of the staging_cfg, staging_RL, RL, no_cfg or rescaled_cfg either as they are just flash variations at its cores. silveroxides/Chroma-LoRAs at main
@qek Those LoRA are reuploaded at a different repository with correct Licensing under them now. There was just a grace period between me unpublishing them and figuring out the logistics.
@S1LV3RC01N I will do some tests when I have time and describe my preferred settings and/or bake in my fave.
Thanks for this, the licensing is part of why I decided to upload this model instead of just keeping my finetune experiments lost and forgotten, like I've done in the past. The community should try to build up the ecosystem of what is - to my knowledge - the only really 'open' model. Would be a shame if I messed that up immediately :-p
@S1LV3RC01N well..switching to your lora GREATLY improved low-step results! Great work, and thanks again!
@dawncreates so I'm a little confused. is the lora re-baked into this flash model? or do I still need to go pull it from S1LV huggingface?
@QualityControl This thread was started about a deleted version. If you download the flash version, it has the rank-128 baked in. That rank seems to give good results with res_2m and bong_tangent at 15 steps and CFG 1. There are probably other rank/sampler/step settings that can work well with even fewer steps - but for that you'd have to get the base model and test different settings and lora ranks.
@dawncreates ok perfect. that's what I figured but didnt see it explicitly stated in the notes and these comments were still so fresh. its a very impressive finetune. Thanks!
Hey, thanks for doing this. However, I'm not getting any of the outputs you shared. I imported your images and ran them with the exact same models, and the results are showing up as if there's not enough steps or CFG. Same seeds and all settings will not give me the same result. Just bad mutations and degraded images.
I should have made this clearer. All example images were made with a low-step lora. I'll try to update the model with non-lowstep examples when I have time. It won't be tonight.
I used the model with the low step burned in. Did you use an additional lora on top of that?
@blastermaster123 No. that should have worked. CivitAI was bugging a bit when I was uploading the lowstep version. At one point there were two different versions of it in the model page, so something might have gotten bugged. I have taken it down anyway. I will try to see if I can get a new lowstep version up, as soon as possible, but it might be a couple of days. With that sampler something like CFG 2.5 and 24-ish steps can give decent results
As a quick test I got good result with rank 64 of https://civitai.com/models/2032955?modelVersionId=2300965 with res_2m/14 step/1cfg settings. Not 1-to-1 but looks good. I will do some more testing when I have time and update the examples and the model description.
@blastermaster123 If you downladed the low-step version I'd strongly recommend deleting it and changing to the base version with a lora, or to the new flash version. It is a massive improvement.
@dawncreates Hate to bring bad news, but it looks like you may have uploaded the wrong flash version again. It doesn't have the flash lora baked in. It requires high steps and CFG to get a decent image.
@blastermaster123 I downloaded and tested it. I use it at 15 steps with res_2m, bong_tangent and CFG: 1. It is the one I meant to upload - but maybe it's not 'flashy' enough? If it's too slow for you I'd get the base version and try some of the other lora-ranks from here: https://civitai.com/models/2032955/chroma-flash-heun If you find a mix and setting you like, feel free to let me know and I can test it and see if a flashier flash should be added.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.