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
Flash version of v1.2. Rank-128 of flash-heun baked in.
FAQ
Comments (29)
Really liked 1.0 (bf16). But 1.2 (bf16) something is very wrong, not good.
For me it is has all the issues of flux with none of the benefits of chroma.
gotta agree. earlier versions were amazing. not pleased with this version at all.
This is interesting. I see issues with both versions but I definitely get better results from 1.2. Could you be a bit more specific about what issues you're having?
@dawncreates
Prompt adherence is degraded to flux levels. Such that it ignores or misinterprets most of the prompt, as one person notes it completely ignore ages. But it is not just ages it is messing up anatomical details. Weird extra clothing bits, like cloth hanging of a sock. Out of proportion hands/feet. One time it generated a giant man compare to the lady, including the furniture.
Prompt:
```
-- pos
Realistic, clear details, 4k. High definition, realistic. A 18 year old girl laying down on the left side of the couch, looking at her phone.
-- neg
sketch, drawing, illustration, painting, art, cartoon, anime, 2d, 2.5d, unreal engine, render, CGI, 3d, rendering, glossy render, stylized lighting, plastic skin, airbrushed, fake, low resolution, distorted, blurry, deformed feet, fused limbs, fused finger, fused toes, blurry background, blur
```
That prompt one time generated the top half of 35 year old lady's torso placed upright on a couch. It also made the room dark to show the phones ambient light. The output was so bad I honestly thought you uploaded the wrong model, I've seen something like this before where someone tried to merge a flux model into chroma (not saying you did this).
I'm sorry I did not spend much time, 4-5 gens then I gave up. I mean 1.2 compared to 1.0 is not even remotely comparable. I have minor issues with 1.0 like missing toes and such.
@zoot_allure855 That's weird! Thanks for the info. Was this with a flash-lora or without? With CFG or without? I've really not had any issues like that at all when I was testing it and I get better quality outputs in 1.2 in my testing - otherwise I wouldn't have uploaded it. But you're clearly not the only one with issues.
There are no loras in this model - it's only my fintune and base Chroma - but they are merged together, so the issues could be from that.
@dawncreates
No loras, just your 16bit model at 40 steps, 3.5 cfg, 16:9 at 768. rk_beta and bong tangent. These are all the same settings used with release 1.0.
@dawncreates Have you compared your local models hash with the one civitai is showing (9AFA766990)?
I confirm the age issue. Without couple cheats it doesn't listen to age tags. It seems that "woman/female" tags recieved extremely high weights and just supress age tags. Prompting of groups of characters also became more complicated comparing with vanilla.
UPD: Added post to gallery.
@mphobbit thanks for the info. I've taken a look at the new captions and they're probably a big part of the age issue. Too much v1.2 in some layers might be anther part of the prompt adherence issue.
I wasn't planning on frequent updates of the model - but I've made some changes and started a new training run. I will update if/when the next version has the best of both versions
@dawncreates Thanks!
Wow. I thought this was an exaggeration... but no. BURN 1.2 WITH FIRE.
Really a big regression, new version is not good at all. Not sure what happened.
This is interesting. I see issues with both versions but I definitely get better results from 1.2. Could you be a bit more specific about what issues you're having?
Yes, version 1 is an amazing model. Version 1.2 is meh outside of the style which can be implemented really easily even if we aren't very talented with prompting. But kudos to the creator, v1 is probably the best model I've tried ever.
version bf16 1.2 stills have problems with characters Lora sadly,a little more than previous version, and now some problems with words,also you put in description to not mention resolution but in your images are stuff like "8k definition"?
But still,the skin looks quite better,less grain finally,and don't have problems going at 4-4.5cfg scale, can't wait for futures uptades
The part about 'not mentioning resolutions' is in the prompt for JoyCaption, so it doesn't include it in the training captions. I don't know how useful it is to put resolution in the prompts, my guess is not very useful, but I don't know. Testing takes a lot of time and I'm a bit time constrained currently, so my settings and prompts should not be considered optimal.
I also find skin and image quality better in 1.2, but people seem to be having issues with this version, so I'll try to do some adjustments and hopefully I can do another training run soon.
@dawncreates ohh I get now, if possible,could you test some compatible with some characters Lora for your next versions?
I'm going to have to agree with the others here, and say that the new version has some major issues. I could easily prompt 20 year old woman in the previous version, but now everybody is 35+ years old for some reason. Did you remove a lot of the dataset?
Flash 1.0 was amazing though, so please keep it up. You're doing great work!
Interesting, I haven't experienced this issue (I've done little testing with flash/flash-lora, so maybe it is less of an issue with CFG..) and I get better results with 1.2 (better skin, clarity, backgrounds). Clearly I need to do more tinkering though. I see some issues with 1.2 that I think could come from being overtrained on a subset of the dataset. I do want to retrace my steps between 1 and 1.2 a little bit and train a new version when I have time, so don't hesitate to comment/message if you have other problems. Would be good to know what problems people have before retraining.
I've also largely found the same to be true. Prompt adherence is completely off with the new version — often it produces completely different people, actions and settings as if something isn't wired correctly. But I've checked the workflow and it's all fine. Plus, when the old version is loaded instead with everything else kept the same, it works fine again. Not sure of the cause, hopefully it can be worked out as this is a really promising model for Chroma!
Good improvement over the original chroma, good job.
I was hoping that 1.2 would remove the massive amount of grain. Instead it ruined everything. Hoping that 1.3+ follows as a branch off 1.0
I'm not sure what this model does better than base Chroma.
Try v1, v2 degraded significantly.
I think it's a very nice model.
It can easily get realistic results without using LoRAs or complex prompt engineering (in my opinion, base Chroma is a bit random in the context of realistic generations). Also, this model has its own style, which is a nice change.
It appears some people say there is something wrong with the new release. I personally only used this model with a Low Step Speed LoRA and haven't used the previous versions or compared them to 1.2, so I can't comment on that. Although, if there are some bugs in the new release, I hope version 1.3 will fix them.
Anyway, thanks for sharing this fine-tune. I know it's a lot of work!
Just to set the record straight dawncreates fine tune version 1.0 is probably the best realistic fine tune for chroma thus far. It is better than 2K-DC, it is better than SG161222's spark preview. I would not critic on version 1.2 unless I thought something was wrong. What I've seen many say is chroma can be fragile, and can easily be over tuned; maybe one day I will fine out for myself. Anyway, if you like dawncreates 1.2 please try 1.0 because it is currently the best (IMO) realistic fine tune for chroma.
I wish dawncreates no ill will, and even if dawncreates does not release 1.3, 1.0 is a fantastic model.
Not a fan of v1.2. Except for the noise, the results were pretty good in v1.0. There is noticeable degradation in 1.2.
v1 is so much better than v2. v2 may have less noise, but it is really bad at following prompt instructions, even with NAG it basically does what it wants. Here's hoping there's a v3 and it has the best of both worlds!
Awesome job on 1.3 flash. In 10-15 test gens, it performs better than 1.0 on 9 out 10 of the prompts. Thank you for the incredible effort and work. Please keep it going!