Edit model card

🚨 IMPORTANT - PLEASE READ EVERYTHING 🚨

This model is an early experimental model of our future planned Flux model. It it NOT representative of a final product and has not undergone the necessary training to be considered so. Many tags have not been included within the model for the purpose of this alpha. You may experience certain problems relating to prompt cohesion, text legibility, species understanding, anatomy, style, and backgrounds. LoRA’s that have been trained on this model may or may not work in future iterations of the model. ANYTHING WITHIN THE MODEL IS SUBJECT TO CHANGE.

Join Our Discord Server!

We have a Discord server where we provide updates on our models, take feedback relating to our models and generally enjoy a good chat. We’d love to see you there!

https://discord.gg/3GZHQTEEJq

Recommended Settings

For optimal results with our model, we recommend the following settings:

  • Resolution: 1024x1024
  • Steps: 15-30
  • CFG: 1 (Disabled) or 2-4
  • Guidance: 4
  • CFG Skip Steps: 4

The choice between CFG 1 and CFG 2-4 depends on your specific needs:

CFG 2-4 advantages:

  • Produces more coherent text, even with lower step counts
  • Can generate more detailed images or results closer to base Flux

CFG 1 advantage:

  • Faster generation times (about half the time of CFG 2-4)

We recommend experimenting with both to find your preference. If using CFG 1, you may need to increase the step count for more coherent results.

When using CFG 2-4, it's crucial to skip steps. We suggest skipping 4 steps at 15 total steps. If your generations appear very blurry and blue, try increasing the number of skipped steps. Our ComfyUI workflow will handle this automatically.

Usage Instructions

Please download the ComfyUI workflow from the link below:

https://huggingface.co/lodestone-horizon/furry-flux-pilot-alpha/blob/main/comfy-workflow.json

  1. Click </> raw
  2. CTRL + S
  3. Save to a location that you know.

Then drag and drop this workflow into your up-to-date ComfyUI interface.

If you do not have ComfyUI, please follow these installation instructions:

https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#installing

CLIP and "UNET" Once you have loaded the workflow, you must select the appropriate CLIP and “UNET” model. The model should be placed in (ComfyUI Folder)/models/UNET and the CLIP model should be placed inside of (ComfyUI Folder)/models/CLIP

VAE: You must also download the VAE for FLUX.1 Dev from here: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors. You do not need to do this if you have already downloaded the FLUX.1 Dev VAE. The VAE should be placed in (ComfyUI Folder)/models/VAE

T5: You must also download T5 from here: https://huggingface.co/comfyanonymous/flux_text_encoders. You do not need to do this if you have already downloaded T5. T5 should be placed in (ComfyUI Folder)/models/CLIP

If you have placed these files after the interface has already been loaded, please refresh the page, otherwise they will not appear.

Both the “UNET” + CLIP was trained with this model, so it’s vital to use both of them for appropriate performance. In the workflow, you should select your model files here.

On the left side of the workflow, you’ll see the settings for your generation. At the top in the green box, you’re able to input your positive prompt. The red box is for negatives.

Please read Prompting & Tips below to learn more about how to prompt the model effectively.

Model Info

Chromafur Alpha is an experimental furry AI model built on Black Forest Labs' FLUX.1-dev foundational model. Created by Horizon Team, it's an initial experiment for a larger model planned in the near future. Chromafur Alpha specialises in generating high-quality SFW and NSFW furry artwork as it has been trained on a focused dataset.

The model uses a custom in-house captioning model designed to describe furry artwork naturally, avoiding overly flowery language. It also incorporates both existing and AI-generated tags, allowing it to respond well to prompts while maintaining Flux's strong natural language understanding.

Notably, Chromafur Alpha has demonstrated the ability to use both tags and captions for image generation, whilst excelling at complicated prompts with extensive natural language and tags.

Strengths & Weaknesses

During our evaluation of the model by ourself and the community, we find the following strengths and weaknesses in the model.

Strengths

  • High-level understanding of natural language, even with complex prompts involving a large variety of objects, fur, colors and more.
  • Flexible natural language style, accommodating complex and simple English..
  • Proficient at creating visually appealing anthropomorphic characters.
  • Capable of generating genitalia and other NSFW elements.
  • Ability to work with various prompt lengths, from concise to detailed.

Weaknesses

  • Limited capability for duo+ compositions, as it's not specifically trained for this.
  • Limited ability to modify style of generation, there tends to be one singular ‘house style’.
  • Image quality tends to degrade with aspect ratios beyond 1:1.
  • Tendency to accidentally include genitalia on characters when not specifically requested.
  • Characters often default to nude unless specific clothing items are prompted.
  • Reduced text generation comprehension compared to base Flux.
  • Tendency to include human figures in backgrounds, even when only furry characters are requested.
  • Occasional issues with image quality, resulting in graininess or blurriness.

Prompting & Tips

It's strongly recommended to use natural language in your generations and to utilise our ComfyUI workflow. Here's how to prompt effectively with our workflow:

ComfyUI Workflow Prompting

1. CLIP Box (Top)

  • Format: image tags: tag1, tag2, tag3
  • Use this for listing image tags, as you would’ve done with SD 1.5 or SDXL.
  1. T5 Box (Bottom)
  • Format: image tags: (same as your CLIP tags) image captions: "your caption here"
  • Include both tags and a natural language caption

Using Negatives

  • Negatives are supported, following the same tagging and prompting guidelines
  • Important: To use negatives, you must use a CFG setting other than CFG 1
  • Negatives will not work with CFG 1

Prompting Style

  • Describe your scene in a natural manner
  • Refer to our provided example images to understand the prompting style
  • Use natural language for detailed descriptions

By following these guidelines, you can effectively leverage our model's capabilities and achieve better results with your generations.

Credits

This model was developed by the Horizon Team, a team dedicated to creating high-quality furry AI models.

  • Lead Research: Lodestone Rock
  • Data: Bananapuncakes
  • Research: Theos
  • Research: Clybius
  • Technical Assistance: Dogarrowtype

Extra Funded Provided By:

We'd like to express our heartfelt gratitude to our wonderful Supporters who made this model possible. Golden Supporters received early access to this model as a thank you for their contributions ❤️

Golden Supporter:

  • 3eve3
  • Gushousekai195
  • Kadah
  • Mobbun
  • Robke223
  • TheGreatSako
  • TheUnamusedFox

Serious Supporter:

  • IlllIs

Supporter

  • degreeze
  • Tails8521
  • Anonymous Donor

If you’d like to support us, you can subscribe to us via SubscribeStar here. A variety of tiers are available.

https://subscribestar.adult/lodestone-rock

We also take donations via Ko-fi.

https://ko-fi.com/lodestonerock

Please contact us if you wish to donate via other means or can provide computational hardware.

Special Thanks

  • IlllIs
  • Mo
  • GodEmperorHydra
  • Furry Diffusion Moderation Team
  • You!

Have a good one! \o/

Downloads last month
350
GGUF
Model size
11.9B params
Architecture
flux

2-bit

3-bit

4-bit

5-bit

6-bit

8-bit

16-bit

Inference API
Unable to determine this model's library. Check the docs .

Model tree for lodestone-horizon/chromafur-alpha

Quantized
(13)
this model
Quantizations
1 model