santa hat
deerdeer nosedeer glow
Sign In

UltraReal Fine-Tune

347
4.4k
153
Verified:
SafeTensor
Type
Checkpoint Trained
Stats
2,595
Reviews
Published
Dec 15, 2024
Base Model
Flux.1 D
Training
Steps: 205,560
Hash
AutoV2
8394D797B3
Bronze Flux Badge
Danrisi's Avatar
Danrisi
The FLUX.1 [dev] Model is licensed by Black Forest Labs. Inc. under the FLUX.1 [dev] Non-Commercial License. Copyright Black Forest Labs. Inc.
IN NO EVENT SHALL BLACK FOREST LABS, INC. BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

What's New in v2.0?

  • Enhanced Anatomy: Hands, feet, and poses have seen major improvements, offering more natural and accurate results. Say goodbye to overly distorted limbs!

  • Improved Textures & Quality: Upgraded skin details, richer textures, and sharper results overall. Blurred images still happen occasionally, but much less frequently than in the previous version or when using LoRAs alone.

  • Improved Text Rendering: Efforts have been made to improve the generation of text in images, and it’s much better than before. However, artifacts can still occur, and strange symbols might sometimes appear instead of readable words. This remains a work in progress.

  • Expanded Dataset: A larger and more diverse dataset (1800 images) introduces better balance across styles, lighting, and compositions.


Added Checkpoint Variations

To ensure compatibility with different workflows, I’ve included multiple checkpoint variations:

  • BF16

  • FP8

  • Quant 8 (Q8)

  • Quant 4 (Q4)
    NF4

From my testing, I’ve noticed Quant 8 (Q8) offers slightly better quality than FP8, providing finer details while maintaining manageable resource requirements, but other works nice too. Pick the version that works best for your setup


Known Limitations

  • NSFW Capabilities: Still a weak area in this version. However, a minor fine-tune focusing specifically on NSFW content is already in the works.

  • Text Rendering: While text generation is better, occasional artifacts like odd symbols or incomplete words may still occur. But noticied usage of t5xxl fp16 instead of fp8 helps a lot with text


Tips for Optimal Results

  • Sampler: Use DPM++ 2M samplers for smooth and consistent outputs.

  • Steps: Aim for 30–50 steps to capture finer details without over-processing.

  • Scheduler: Beta Scheduler remains the best choice for this checkpoint.

    Prompting Tips

    The best prompting style involves complex prompts with clear, comma-separated phrases. While you can get creative with storytelling prompts, unnecessary descriptions like “this crap added more vintage to her style” won’t improve the results. Keep it concise and descriptive, focusing on essential visual details for the best output.


Future Plans

I’m committed to further developing this fine-tune. The next update will likely focus on:

  • Expanding NSFW capabilities

  • Enhancing edge cases like dynamic poses and lighting scenarios

  • Improving text rendering for sharper, more accurate results

    P.S: If you still don't have realistic effect, then try add my ultrareal lora, usually helps me a lot




    Ultra-Realistic Flux Fine-Tune v1

This is my first experiment in fine-tuning a checkpoint, built upon the foundations of my UltraReal LoRA and expanded with an extended dataset. The aim? To push realism to the next level, finding that sweet spot between amateur aesthetics and professional, high-quality visuals.

While this is only the first version and I see room for further refinement - the results are good, but not ideal (hands and feet can be broken sometimes, but situation is not critical, still better then defaul flux). This fine-tune isn’t just about amateur-quality outputs; it shines with professional-grade images, offering exceptional detail, lifelike shadows, and lighting. It’s a versatile model designed to unlock a wider range of realistic image generation possibilities.

This is very much a work in progress, and I’m sharing it to gather feedback and see how others use it creatively. If you test it, I’d love to hear your thoughts or see your results!
Also i uploaded both versions: fp16 (in ComfyUI it's better to use with e5m2) and fp8 and Q4_0


🌟 What’s New in This Fine-Tune?

  • Expanded Dataset: Nearly double the dataset size of the original LoRA, covering a diverse range of styles, lighting, and compositions.

  • Improved Realism: Sharper details, richer textures, and more natural lighting, bridging the gap between AI-generated and real-world imagery.

  • Versatility: From casual amateur-style snapshots to cinematic, professional-quality renders, this fine-tune adapts to a variety of creative needs.

  • Enhanced Anatomy: Better hands, limbs, and more natural poses compared to the base Flux model.


💡 Tips for Best Results

  • Use DPM++ 2M samplers for smooth and consistent outputs.

  • Aim for 30–50 steps for finer details without overdoing it.

  • Select the Beta Scheduler for optimal rendering performance.


Why Fine-Tune?

This fine-tune was crafted to overcome some of the limitations of the default Flux model. It enhances its ability to handle complex scenes while maintaining consistent quality across a range of prompts. The goal is simple: make ultra-realistic image generation accessible, reliable, and visually stunning, without requiring endless adjustments.

P.S: i plan to train this model more to make ultimate checkpoint with best anatomy and realism. This version is not very good with nsfw (this will be fixed in next version)
P.S.S: so far you can randomly get a low resolution image (dunno what exactly trigger this one, but will search for fixes). But seems like using high-resolution in prompt helps