CivArchive
    Preview 124312333
    Preview 124312342
    Preview 124312326
    Preview 124230473
    Preview 124312336
    Preview 124230477

    Azure Starfire

    Azure Starfire is a furry-focused base model, designed to deliver strong anatomy, smooth texture detailing, and high visual consistency across species types.
    Built using selective merging for optimal tone, composition, and fur rendering fidelity

    Starfire Void Stable V5 Update

    🔧 New Merge Adjustments

    • Added Hassaku XL (Illustrious) v3.4 at 0.3 strength
      → Improves structure stability and overall composition balance.

    🎨 Background Stability Fix

    • Integrated Azure Light – GothicLL LoRA at 0.2 strength
      → Helps reduce messy or overdrawn backgrounds
      → Produces cleaner, more readable scenes

    🌑 Visual Direction

    V5 is more subtle and controlled compared to V4

    • Slightly less semi-realistic

    • Maintains darker tonal style similar to V4

    • Cleaner rendering with less visual clutter

    🧠 Prompt Detection Adjustment

    • Improved prompt interpretation for more precise results

    • Better response to structured prompts and LoRA stacks

    Overall Changes

    • Less realism, more stylized output

    • Reduced artifacts

    • Cleaner background handling

    • More stable generation

    • Slightly more restrictive but consistent results

    Stellar Stable V2 Update

    Merge Adjustment

    Experimental Merge Method

    This version uses a new merge block method currently under testing Results may or may not provide overall improvement.Feedback and testing results are welcome as this approach is still experimental.

    Updated merge block ratio to improve balance between detail and noise control

    • Stellar V1 → 0.65

    • Indigo Furry Mix XL – NoobAI_EPS_X_10 → 0.35

    Stellar Stable Update

    Stability Improvements

    • Added Indigo Furry Mix XL – NoobAI_EPS_X_10 to further stabilize background noise and reduce unwanted artifacts.

    VAE Update

    • Swapped VAE to neptuniaXLILNAIVAE_contrastColors for improved color contrast and softer overall rendering.


    Starfire Void Stable V2 Update

    Starfire Void Stable V2 is an upgraded fine-tuned build of the Starfire Void line, designed to deliver greater image consistency, depth, and clarity.
    This version utilizes Azure ColdSnap for the fine-tuning process and introduces a new VAE,

    ⚙️ Technical Details

    • Base: FirstLight Base

    • Fine-Tuned With: Azure ColdSnap

    • VAE: Replaced with a darker variant to improve exposure control

    • Additional Merge: Azure LowLight at 0.4 strength

    • Dataset Used: E6 dataset

    Changes & Improvements

    • Significantly improved generation stability compared to V1.

    • More expressive lighting and refined color depth.

    • Humanoid-based training data used instead of furry, resulting in better balance for mixed-scene compositions.

    • Enhanced background clarity and reduced unwanted detail clutter.


    Dark First Light Stable Changelog

    Dark First Light Stable is a Fine-tuned and stabilized version of the First Light Base, merged with Azure GothicLL Style at 0.5 strength.

    Summary

    Dark First Light Stable delivers a more polished, atmospheric look suited for darker or low-light compositions while maintaining strong consistency , adaptability and controlled with reduced prompt interference.

    Lighting & Visual Adjustments

    • Overall darker and moodier lighting compared to the base version.

    • Improved texture fidelity and clarity in shaded environments.

    Model Improvements

    • Merged with Azure GothicLL to reinforce stylistic shading and surface detail.

    • Significantly reduced unwanted traits in prompt crafting.

    • Produces cleaner, sharper compositions with consistent anatomy.

    • Improved background rendering stability and reduced noise.


    Starfire Void Stable Update Changelog

    🧩 Overview

    Starfire Void Stable is a fully fine-tuned version of the First Light Base, featuring major stability and quality improvements.

    🖐️ Fixes & Enhancements

    • Integrated hand and finger fix for more natural anatomy results.

    • Significantly reduced background clutter and unwanted details, improving focus on the main subject.

    ⚙️ Fine-Tuning Process

    This version was fine-tuned using Indigo Furry Mix XL Esp 7, resulting in noticeably smoother rendering and balanced detail


    🌟 First Light Tuned Change Log

    Overview
    First Light Tuned is a light-tuned variant of the First Light Base Model, developed to enhance detail consistency and improve rendering quality.

    🧩 Dataset Integration
    This version includes additional R34 dataset integration for expanded visual variety and improved adaptability across different styles.


    Base Components Merge

    • YiffyMix

    • Wild Safari

    • MALE-ggmix-Illustrious

    This combination provides balanced color tone, anatomy accuracy, and general versatility, ideal for both anthro and semi-feral forms.

    • Steps: 25–40

    • CFG Scale: 4–8

      • ⚠️ Values above 9 may cause style blending or loss of detail, depending on prompt structure.

    • Sampler: Euler a / DPM++ 2M Karras (user preference recommended)

    • Hires Fix: ✅ Recommended

    • FaceDetailer (ComfyUI): Optional, best used for close-up portraits

    🏷️ Prompt Tags

    ✅ Positive Tags (Quality Boost)

    masterpiece, amazing quality, best quality, ultra-detailed, 8K, illustration, CG, ultra-detailed-eyes, detailed background, very awa, newest, highres, absurdres

    🚫 Negative Tags (Suggested)

    lowres, (worst quality:1.2), (bad quality:1.2), (low quality:1.2), bad anatomy, old , mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers

    Description

    🌟 Stellar Stable V3 Change Log

    Stellar Stable V3 is intended as a maintenance update, keeping the same visual style and merge structure as V2 while addressing issues found in the previous release.

    🔧 Fix Update

    • Merge composition remains unchanged from Stellar Stable V2.

    • This release focuses on fixing issues present in V2, with no major structural changes to the model.

    🛠 Improvements

    • Resolved stability problems reported in V2.

    • Improved overall generation reliability.

    FAQ

    Comments (12)

    Crovidius7849Apr 14, 2026
    CivitAI

    Hey, nice model, looking good so far. How well can it handle images with 2 character?

    BastionExa
    Author
    Apr 14, 2026

    From my current test V3 Stellar should be more stable generation for more than 1 or furry interacial

    Crovidius7849Apr 14, 2026

    @BastionExa Thanks, I'll give it a shot. Does your model also use danbooru style tagging? From what I saw it does, but I want to make sure before diving into this experiment.

    BastionExa
    Author
    Apr 14, 2026

    @Crovidius7849 Should be Based on Danbo dataset

    Crovidius7849Apr 14, 2026

    @BastionExa Cool, I'll run some tests. Previously I experimented with Yiffinhell and Novafurry, but even with an IPAdapter, the results were mostly mixed (prompt bleeding all over), rather than consistent (character design wise).

    BastionExa
    Author
    Apr 14, 2026

    @Crovidius7849 i think that is the current weakness SDXL in general, multiple character

    Crovidius7849Apr 14, 2026

    @BastionExa Agree, although seeing others manage this composition (not just spicy), I also want to figure out a way to do it. Even knowing that it must've required them many generations to have 2-3 decent image.

    BastionExa
    Author
    Apr 14, 2026

    @Crovidius7849 regional prompting or LoRa of X character should help alot with my experience for multiple char

    Ashen_1Apr 14, 2026

    @Crovidius7849 With just txt2img you can get consistent results by using contrasting characters and keeping descriptions to a minimum. If you have two characters of the same or similar species then their details will get mixed up. Character loras that use 1-2 keywords should work but I haven't personally tested it. Anything else will need to be done with some form of inpainting, be it regional prompting as mentioned or manually drawing the masks.

    Crovidius7849Apr 14, 2026

    @Ashen_1 Thanks your response, in my previous attempt with a different model (Yiffinhell), I also approached it this way, lion-dragon, lion-fox and lion-gryphon pairing was that I tried (different scenarios), but many times have gotten mixed results (such as gender, color, features, complete gender-species pair mixing and so on). With this I need to test these out. We'll see. But I agree, the simplistic tagging helps (using Claude to engineer the prompts, and we figured this as well).

    Ashen_1Apr 14, 2026

    @Crovidius7849 I've taken a look at your duo images and here is some advice I can give you.
    Use whitespace for tags instead of underscore so for example "looking at viewer" instead of "looking_at_viewer"
    Never use "BREAK" unless Regional prompter uses it to differentiate between parts of the image. I can't explain it properly but it pads the prompt with empty characters until the token block is filled. When you have a lot of tokens then parts of your prompt will get ignored. Different blocks do not have different context though so you can't use it to seperate one character from another or the background.
    Tag order itself shouldn't matter but you do get differences when you swap the order around so it might be worth experimenting with that. I tend to put characters first and background last but thats mostly to keep it organized.

    Crovidius7849Apr 14, 2026

    @Ashen_1 Thanks again, but if I remember well, the recent duo images I tried to generate are not uploaded here, and those you can see in my gallery were generated using ForgeUI. Recently switched to Comfyui and adjusted a workflow to my needs, but if it helps I can share one of the better duo images for you to check.

    Checkpoint
    Illustrious

    Details

    Downloads
    888
    Platform
    CivitAI
    Platform Status
    Available
    Created
    3/16/2026
    Updated
    8/5/2026
    Deleted
    -

    Files

    azureStarfire_stellarStableV3.safetensors