NOTE: The original v1 had some bugged layer names - this is now fixed (as of the evening of the 31st of October). Having the wrong version shouldn't affect generation in ComfyUI - but it might affect things like training and quantization.

A photorealism-focused finetune of Chroma (https://civitai.com/models/1330309/chroma), designed to make producing realistic, photo-style images easier and more consistent. All example images were made using some sort of flash or low-step lora.

Chroma is a fantastic and highly versatile model capable of producing photo-like results, but in my experience it can require careful prompting, trial-and-error, and/or loras. This finetune aims to improve reliability in realistic/photo-based styles while preserving Chroma’s broad concept knowledge (subjects, objects, scenes, etc.). The goal is to adjust style without reducing other capabilities. In short, Chroma can probably do anything this model can, but this one is more lenient.

Personally I'd recommend downloading the non-flash model, then you can experiment with steps and CFG and choose which flash-lora best suits your needs (if you need one). The flash version of the model has the rank-128 lora baked in (https://civitai.com/models/2032955/chroma-flash-heun).

Example Generation Notes
The example prompts are not meant to be examples of good prompts - several are copy/pasted from other images on CivitAi . As a note, tags seem to push the model closer towards art/anime, natural language towards photos. For photos use terms like photo or photography, but avoid photorealistic. Photorealistic is good for photorealistic art. In general, simply describing what you want to see in natural sentences seems to work really well.

I imagine any of your current Chroma settings will work well. I used the following generation settings for the example images (these are not necessarily optimal settings):

• Workflow: Basic Chroma workflow in ComfyUI

• Some flash LoRA: Strength 1

• Megapixels: 1 - 1.5

• Steps: 14-15

• CFG: 1

• Sampler: res_2m

• Scheduler: bong_tangent

All example images were generated without upscaling, inpainting, style LoRAs, subject LoRAs, ControlNets, etc. Only the most basic t2i workflow was used.

Training Details
The model was trained locally on a medium sized collection of openly licensed images and my own photos, 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. The database consists almost exclusively of SFW-images of people and landscapes, so to retain Chroma-HD's original conceptual understanding, selected layers were merged back at various ratios. All the juice comes from Chroma itself (get to work finetuners)!

I aim to continue finetuning and experimenting, but the current version has some juice.

All images were captioned using JoyCaption:
https://github.com/fpgaminer/joycaption

The model was trained using OneTrainer:
https://github.com/Nerogar/OneTrainer

Support
Have too much money? Want to support me?
https://ko-fi.com/dawncreates