Boring Reality LoRA for Flux Dev 2
This LoRA is an early experimental training attempt at Flux Dev 2. The results are not perfect, but they should have at least limited improvement on Flux generations (distortion and disfigurement will increase with issues like text getting worse.)
A noticeable impact on this Lora is not just that it increases the "realism" of the images but that they tend to have better world knowledge and can produce better results in other styles such as cinematic shots and animation.
You will usually have to increase the guidance by a lot when raising the scale of the LoRA for better results. More descriptions in the prompt tend to help as well.
There is still a lot of potential on Flux Dev 2 from the creativity side of each generation that other recent models have not been able to achieve.
NOTE: The ComfyUI version may have not been converted properly and may now perform as well as the Diffusers version. There may need to be better improvements in the workflow provide and what not as well.
Description
FAQ
Comments (15)
oh wow. i see your post, i like haha. i can see that you left your comfort zone and really went full super saiyan mode on this one :)
Yea, I did not do anything different with the dataset really. I just tried to have images for this model stand out a bit more as Z Image took up all the attention. Flux Dev 2 still has a lot of potential creativity wise that nobody seems to have tried out yet. I had not even had enough time on it while focusing on training runs for Z image as it is.
@kudzueye for a second i thought you accidentally uploaded random stuff from your Pictures folder lol. good work, boss
@kudzueye 🙏👏🙌 Yeah it looks like F2 V Z Image might split the men and the boys. 😉 And the boys are always louder. I'm all about Flux these days and thrilled with how far I've gone with Flux(1) Dev - all thanks to LoRA. It's sad to see so much Z and so little F2 on Reddit. I'm thrilled that at least some people are giving it attention.
Awesome. Thanks a lot. Can you share your training config (rank 32,64?) and number of images ?
The best training run was done on a provided service where I had almost no control on anything such as rank. You can probably deduct the rank though from the file size if you have any existing Flux2 loras from AI Toolkit or what not by comparing 2x multiplier factor on files by their rank.
Steps was at 2000. Learning rate I believe at 0.00005. Around 30 images trained on. I cannot remember if I used actual captions on the run for that or just a general "photo" caption for everything. The ComfyUI version uploaded here was quickly converted from the PEFT format with some vibe coded script that may have not translated everything all that well in case it under performs.
I will have a better idea with more specifics on training config, when I have time to rerun and test the datasets in AI Toolkit though at the moment Z Image is taking all the time there.
where can I find a workflow for loras on flux.2
You can start out just using this one: https://huggingface.co/kudzueye/boreal-flux-dev2/blob/main/flux_dev_basic_lora_workflow.json
absolute cinema
Very good first attempt! Results are satisfying. Thanks for the tip about increasing Flux Guidance to 10. I was wondering why my first try looked really bad. Is this a quirk with the new Flux.2 or is it due to the way the LoRA was trained? I don't remember having to do this with Flux.1.
I am not really sure myself, but the LoRA might be partially over trained in areas such that it will break apart and the higher guidance gets it back to the original prompt. Surprisingly unlike other models, it does not over bake the image with the very high guidance when used with the LoRA.
incredible!!!
you plan launch to z-image to?
nice one ... if you're not already doing it ... a hint train with json format like
```
{
"scene": "a close-up portrait of a young woman smiling warmly at the camera, set against a textured red brick wall background.",
"subjects": [
{
"description": "a young caucasian woman with fair skin and light brown hair styled loosely around her shoulders. she has striking green eyes and is wearing large, dangling silver earrings with feather accents and a dark red top. her sunglasses are perched on top of her head.",
"age": "20s",
"hair": "light brown, wavy, partially pulled back",
"expression": "warm, genuine smile showing teeth",
"emotion": "happy, confident, approachable",
"position": "centered in the frame, filling most of the composition from chin to shoulders",
"action": "looking directly at the camera with a bright, engaging expression"
}
],
"style": "professional portrait photography with a clean, slightly stylized aesthetic.",
"color_palette": ["#ff4500", "#8b0000", "#f5deb3", "#ffffff"],
"lighting": "soft, diffused natural light illuminating her face evenly from the front, creating gentle highlights and minimal harsh shadows. the lighting is flattering and enhances her features.",
"mood": "positive, cheerful, relaxed, and inviting",
"surroundings": "a weathered red brick wall with visible mortar lines serves as the background, adding a rustic and urban texture to the image.",
"composition": "centered portrait composition focusing on the subject's face. the framing is tight enough to emphasize her expression while still including some of her shoulders and hair for context.",
"camera": {
"angle": "eye-level frontal shot",
"lens": "standard prime lens \(e.g., 50mm\)",
"depth_of_field": "shallow depth of field, keeping the subject's face in sharp focus while softly blurring the brick wall background"
}
}
```
After some testing in ComfyUI I found that using this model at higher strength tends to introduce a lot of deformities and anomalies, but using the FlowMatch Euler Discrete Scheduler custom node (https://github.com/erosDiffusion/ComfyUI-EulerDiscreteScheduler) proves to work well in reducing those faults while still keeping the Lora's style. It's also the scheduler that the official Flux 2 diffuser uses in their config which isn't currently exposed natively in ComfyUI (https://huggingface.co/black-forest-labs/FLUX.2-dev/blob/main/scheduler/scheduler_config.json) - just don't enable dynamic shifting in the node
想问一下您的FlowMatch Euler 离散调度器自定义节点的参数配置，您指的更高强度是1.0吗？还是更高，想问下，lora强度和法库县引导的强度您是如何设置的？