base_model: prithivMLmods/Qwen3-VL-8B-Instruct-c_abliterated-v3
datasets:
- hearmeneigh/e621-rising-v3-small
pipeline_tag: image-text-to-text
library_name: gguf
license: apache-2.0
tags: - not-for-all-audiences
- gguf
- qwen3-vl
- vision-language
- furry
- scalie
- dragons
- booru
- e621
- danbooru
- abliterated
- uncensored
language: - en
- zh
Qwen3-VL-8B-Instruct-c_abliterated-v3-furry
A domain-specialised vision-language model for Dragon, Furry, and Scalie artwork — built for
high-throughput booru-style auto-tagging and high-fidelity natural-language image captioning.
This is a QLoRA fine-tune of prithivMLmods/Qwen3-VL-8B-Instruct-c_abliterated-v3,
which in turn derives from Qwen3-VL-8B-Instruct. The fine-tune concentrates the model's visual
attention on anthropomorphic, draconic, and reptilian characters, and aligns its output vocabulary
with the e621 and Danbooru tag standards. The base model's abliterated (refusal-suppressed)
behaviour is inherited and preserved.
⚠️ Content Warning
This repository contains a model that is explicitly designed to describe NSFW / Explicit artwork.
- The training corpus contains explicit adult (NSFW) e621 and Danbooru tag annotations, and the
model is capable of producing detailed descriptions of adult sexual content. - The base model is an abliterated checkpoint: it has had its refusal behaviour suppressed and
will not decline explicit requests, and it may produce unprompted explicit content. - The repository is tagged
not-for-all-audienceson the Hugging Face Hub for this reason.
Use this model only for lawful artistic research and personal dataset annotation. You are
responsible for complying with the laws and platform rules that apply to you. Do not use it to
produce, distribute, or facilitate content that is illegal where you are, that involves minors, or
that infringes on the rights of others. By downloading this repository you accept full
responsibility for how you use the model. See Disclaimer below.
Model at a Glance
| Property | Value |
|---|---|
| Model name | Qwen3-VL-8B-Instruct-c_abliterated-v3-furry |
| Base model | prithivMLmods/Qwen3-VL-8B-Instruct-c_abliterated-v3 |
| Architecture | Qwen3VLForConditionalGeneration (text: 36 layers, hidden 4096; vision tower: depth 27) |
| Parameter count | ~8B (Qwen3-VL dense) |
| Distribution format | GGUF (llama.cpp) |
| Main weights | Q8_0 quantisation — 8.11 GB |
| Vision projector | mmproj, f16 — 1.08 GB |
| Context length | 4096 – 32768 (runtime dependent; 8192 recommended) |
| Modalities | Image → Text (single image, multi-turn chat supported) |
| Primary languages | English (en), Chinese (zh) |
| Tag standards | e621 (General Top 10k), Danbooru |
| Base-model traits | Abliterated / refusal-suppressed |
Files in this Repository
| File | Role | Size |
|---|---|---|
Qwen3-VL-8B-Instruct-c_abliterated-v3-furry-Q8_0.gguf |
Main model weights (LLM text backbone), Q8_0 quantised | 8.11 GB |
mmproj-Qwen3-VL-8B-Instruct-c_abliterated-v3-furry-f16.gguf |
Vision encoder & multimodal projector (mmproj), f16 — encodes images and maps visual features into the LLM embedding space |
1.08 GB |
⚠️ Both files are required
The main GGUF and the
mmprojGGUF must be loaded together.
If you load only the main model, it will run as a text-only model and will not accept or
"see" images at all — it will either error out or hallucinate a caption for a blank input.
There is no image understanding withoutmmproj.
The table above lists the exact filenames shipped in this repository. If you rename either file on
upload, update every -m / --mmproj path in the commands below (and in LM Studio) accordingly.
Capabilities
- Booru-style tag prediction — returns canonical
snake_casee621 / Danbooru tags, ordered
from most to least salient. Suitable for producing training captions for SDXL, Flux, Illustrious,
Pony and similar image models. - Structured three-section analysis — with the native prompt, the model emits a
visual feature summary, a canonical tag set, and related/style tags in one pass. - Detailed natural-language captioning — long-form, sectioned English descriptions covering
species, coat/scale colour and markings, pose, anatomy, clothing, expression, composition,
lighting and background. - Tag-space reasoning — e621 taxonomy questions (tag normalisation, aliasing, hypernym
relations, related-tag suggestions) are covered by the text-only portion of the training mix. - No refusal path — inherited from the abliterated base; explicit content is described rather
than declined.
Native Output Format
When prompted with the native training instruction (e.g., "请分析这张图片,输出作品视觉特征总结、标准 e621 通用标签以及同风格关联推荐标签。" or "Analyze this image and provide a visual feature summary, standard e621 general tags, and related style tags."), the model produces three structured sections:
### 作品视觉特征总结
该作品包含以下核心视觉元素:`anthro`(拟人兽人), `areola`(乳晕), `bound`(被束缚),
`breasts`(乳房), `cloven_hooves`(偶蹄), `female`(女性), `hooves`(蹄), `nude`(裸体),
`solo`(单人), `white_background`(纯白背景)。
### 标准 e621 通用标签
`anthro areola bound breasts cloven_hooves female hooves nude solo white_background`
### 同风格关联推荐标签
`restrained intersex`
The section headers are part of the fine-tuned output contract and are emitted in the language of
the instruction. English instructions produce English headers and English prose; the tag tokens
themselves are always canonical English snake_case, regardless of the instruction language.
How to Use
LM Studio
Place both files in the same model folder under your LM Studio models directory, e.g.
LM Studio/models/Qwen3-VL-8B-Furry/:LM Studio/models/Qwen3-VL-8B-Furry/ ├── Qwen3-VL-8B-Instruct-c_abliterated-v3-furry-Q8_0.gguf └── mmproj-Qwen3-VL-8B-Instruct-c_abliterated-v3-furry-f16.ggufSelect the main model. In LM Studio's model search, load
Qwen3-VL-8B-Instruct-c_abliterated-v3-furry-Q8_0.gguf(themmprojfile is not a chat model — do not load it
as one). LM Studio detects the matching Qwen3-VL projector automatically for models in the
correct folder layout.Verify the projector is attached. Open the model's Vision / Multimodal section, or use
the paperclip / image button in the chat composer to attachmmproj-Qwen3-VL-8B-Instruct-c_abliterated-v3-furry-f16.ggufmanually if auto-detection did not pick it up. If the
projector is missing, the image attachment will fail or the model will answer as if it saw
nothing.Attach an image in the chat window (drag-and-drop or the image button), then prompt.
Sampling — recommended starting values:
Parameter Recommended value Notes Temperature 0.4 – 0.7 0.4for reproducible tagging,0.6–0.7for varied captionsTop-P 0.9 – 0.95 Min-P 0.05 Repeat penalty 1.15 Important — see the repetition note below Context length 8192 Raise for long multi-image sessions Max output tokens 512 – 1024 256–512 is enough for a pure tag list Local server (optional). LM Studio's Developer → Start Server exposes an
OpenAI-compatible endpoint onhttp://localhost:1234/v1; use theQwen3-VL-8B-Instruct-c_abliterated-v3-furry-Q8_0.ggufmodel id and send images asimage_urldata URIs.
llama.cpp
Requires a recent llama.cpp build with Qwen3-VL multimodal (mtmd) support.
One-shot image tagging with llama-mtmd-cli:
llama-mtmd-cli \
-m ./Qwen3-VL-8B-Instruct-c_abliterated-v3-furry-Q8_0.gguf \
--mmproj ./mmproj-Qwen3-VL-8B-Instruct-c_abliterated-v3-furry-f16.gguf \
--image ./input.jpg \
--jinja \
-c 4096 -n 256 \
--temp 0.4 --top-p 0.9 --min-p 0.05 --repeat-penalty 1.15 \
-sys "You are a professional booru tagging assistant. You output canonical e621 and Danbooru tag names only." \
-p "List the tags for this image as a single comma-separated list, ordered from most to least salient. Output tags only."
OpenAI-compatible server:
llama-server \
-m ./Qwen3-VL-8B-Instruct-c_abliterated-v3-furry-Q8_0.gguf \
--mmproj ./mmproj-Qwen3-VL-8B-Instruct-c_abliterated-v3-furry-f16.gguf \
--host 127.0.0.1 --port 8080 \
-ngl 99 -c 8192 \
--temp 0.6 --top-p 0.95 --repeat-penalty 1.15 \
--jinja
Then call it with an image:
curl [http://127.0.0.1:8080/v1/chat/completions](http://127.0.0.1:8080/v1/chat/completions) \
-H "Content-Type: application/json" \
-d '{
"messages": [{
"role": "user",
"content": [
{"type": "text", "text": "List the tags for this image as a single comma-separated list. Output tags only."},
{"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,<BASE64_IMAGE>"}}
]
}],
"temperature": 0.4,
"max_tokens": 256
}'
Note —
--image-min-tokens.llama.cppprints a warning for Qwen-VL models that they
generally want at least 1024 image tokens for accurate grounding. This model was trained at
512×512 (image_max_pixels = 262144), so the default tokenisation already matches the training
distribution and no override is required. If you feed high-resolution images and find the tags
too coarse, add--image-min-tokens 1024to upsample the image patch grid.
Prompt Examples
The prompts below were run against this exact build; the shown outputs are real model output.
1 — Native three-section analysis (Chinese / English)
User (ZH): 请分析这张图片,输出作品视觉特征总结、标准 e621 通用标签以及同风格关联推荐标签。
User (EN): Analyze this image and output a visual feature summary, standard e621 general tags, and related style tags.
(See Native Output Format above for the exact three-section response structure.)
2 — Comma-separated e621 / Danbooru tag list
System: You are a professional booru tagging assistant. You output canonical e621 and
Danbooru tag names only.
User: List the tags for this image as a single comma-separated list, ordered from most to
least salient. Output tags only.
anthro, breasts, crossdressing, female, genitals, hooves, nipples, nude, solo, spread_legs,
spreading, suspension, upside_down, vulva
3 — Detailed natural-language description
System: You are a professional Furry/Dragon visual analysis assistant. You describe artworks
in precise, objective natural language.
User: Write a detailed description of this artwork: species, coat or scale colors and
markings, pose, anatomy, clothing and accessories, expression, composition, lighting
and background.
**Pose and Anatomy:**
She is lying on her back, with her legs raised in the air. Her legs are spread wide, and her
feet are raised off the ground. Her hind legs are spread further apart than her forelegs. Her
arms are bent at the elbows, with her hands resting on the ground. Her breasts are prominent,
and her vulva is visible between her legs.
**Clothing and Accessories:**
She is wearing thigh highs and a G-string around her waist.
**Expression:**
The character has a smile on her face, and her eyes are looking at the viewer.
**Composition:**
The character is positioned in the center of the image, and the image is framed with a black
border. The artist's signature "ECM" is visible in the bottom left corner.
**Lighting and Background:**
The lighting in the image is bright, and the background is a simple light color. The image has
a heart symbol above the character's head.
Sampling Note — Repetition Loops
Under a bare tag-only prompt with no repetition penalty and temperature ≥ 0.6, this model can
fall into a degenerate token loop (observed: a runaway y_another_*_focus repetition after a
valid prefix). This is a sampling artefact of the Q8_0 quantisation, not a capability limit.
Mitigations, in order of effectiveness:
- Set
--repeat-penalty 1.15(verified to eliminate the loop attemp 0.4). - Use temperature ≤ 0.4 for tag extraction (verified: clean output at
temp 0andtemp 0.4). - Prefer the native three-section prompt (see above) for bulk work — its structured format
keeps generation anchored and did not reproduce the loop in testing. - Cap
-nto the expected tag count (e.g.-n 256) so a loop cannot run long.
Training Details
Method
| Item | Value |
|---|---|
| Method | QLoRA (4-bit quantised base + LoRA adapters), bitsandbytes NF4 |
| Compute dtype | bfloat16 |
| LoRA rank / alpha / dropout | r = 16, α = 32, dropout = 0.05 |
| LoRA target modules | All linear projections (q/k/v/o_proj, gate/up/down_proj, linear_fc1/2) and mm_projector |
| Adapted modules | 254 |
| Trainable parameters | 43,933,696 (~0.52 % of the model) |
| Vision tower | Frozen |
| Max sequence length | 2048 |
image_max_pixels |
262144 (512 × 512) |
| Per-device batch size | 1 |
| Gradient accumulation | 8 (effective batch 8) |
| Learning rate | 1e-4, cosine decay, warmup ratio 0.08 |
| Epochs | 2.5 |
| Optimizer steps | 4235 |
| Hardware | 1 × RTX 3080 Laptop (16 GB VRAM), peak VRAM 15.78 GiB (torch max allocated 9.81 GiB) |
Training Data
Derived from hearmeneigh/e621-rising-v3-small,
annotated against the Danbooru and e621 tag standards. The fine-tuning mixture combines:
- 10,822 text-only tag-QA pairs covering e621 taxonomy reasoning — tag normalisation, alias
and misspelling correction, hypernym relations, and related-tag recommendation, all resolved
against the canonical e621 General Top 10k vocabulary (10,000 tags). - 3,000 image-caption triples at 512 × 512, pairing artwork with its visual-feature summary,
canonical tag set, and style-associated tags.
The corpus contains explicit (NSFW) annotations, and the model is explicitly trained to
describe and tag them.
Training Metrics
| Metric | Value |
|---|---|
Final train_loss |
0.1384 |
Final eval_loss |
0.2607 |
eval_loss trajectory |
0.3184 → 0.2843 → 0.2602 → 0.2608 → 0.2607 (steps 1000 / 2000 / 3000 / 4000 / 4235) |
| Peak VRAM (nvidia-smi) | 16,161 MB (15.78 GiB) |
Validation loss decreases monotonically through step 3000 and then plateaus through the final
evaluation — a converged solution with no sign of over-fitting.
Evaluation
The released GGUF artefacts were validated at build time:
Structural validation (gguf metadata inspection)
- Main model —
gguf_version=3, 399 tensors, 34 KV entries,general.architecture = qwen3vl,general.file_type = 7(Q8_0), chat template and tokenizer vocabulary present. - Projector —
gguf_version=3, 352 tensors, 24 KV entries,general.architecture = clip,general.file_type = 1(F16),clip.projector_type = qwen3vl_merger.
Content validation (generated text assertions)
- Chinese visual summary ≥ 50 characters, present, and Chinese-dominant within the summary
section. - 89 canonical tags from the e621 vocabulary matched in a single response.
- Three-section structure detected; no refusal or censorship phrasing emitted.
Multi-image spot checks (4 images) confirmed consistent behaviour across subjects.
Intended Use
- Automatic captioning / tagging of artwork datasets for generative-model training
(SDXL, Flux, Illustrious, Pony, and similar). - Bulk normalization of legacy or community-made tags into canonical e621 / Danbooru form.
- High-detail natural-language reverse-prompting and alt-text generation for furry, scalie and
dragon artwork. - Research into tag-vocabulary grounding and multimodal alignment.
Out of Scope
- Not a general-purpose assistant, chat companion, or reasoning model — it is narrowly specialised
for visual tagging and captioning of this artwork domain. - Not intended for photorealistic or non-furry imagery.
- Not suitable for safety-critical, medical, legal, or surveillance applications.
- Not intended for generating images.
Limitations
- Specialisation trade-off. The fine-tune trades broad instruction-following generality for
domain accuracy; the model is heavily biased toward the three-section tag/caption format. - Tag-list hallucination. Long tag lists may contain plausible-but-wrong or over-specified
entries. Sanity-check against the e621 vocabulary, and consider validating every emitted tag. - Quantisation loss. Q8_0 is a lossy 8-bit quantisation. Compared with the BF16 original,
rare-token fidelity and occasionally tag precision degrade slightly. Use the BF16 safetensors
base for archival or evaluation-critical work. - Image resolution. Training used 512 × 512 inputs; very large images may yield
over-confident or incomplete tag sets without--image-min-tokens 1024. - Repetition loops. See the sampling note.
- No safety layer. There is no refusal mechanism by design.
Disclaimer & Legal Use
This model is provided for legitimate artistic research, dataset annotation, and personal
non-commercial experimentation.
By downloading, using, or redistributing this model you explicitly acknowledge and agree that:
- The model is not filtered, moderated, or safeguarded in any way, and it is capable of
producing explicit sexual content describing adult subjects. - You will not use this model to create, distribute, or facilitate any content that is
illegal in your jurisdiction, that depicts or involves minors or age-indeterminate subjects in
a sexual context, that infringes copyright or other intellectual property rights, or that
violates the terms of service of any platform you upload to. - You accept full and sole responsibility for all output produced by this model and for any
downstream use of it. - The authors and contributors of this repository disclaim all liability for any damages,
claims, or consequences arising from your use or misuse of the model, to the maximum extent
permitted by law.
This repository is not affiliated with, endorsed by, or sponsored by the Hugging Face Hub,
Qwen, Alibaba Cloud, the e621 project, or Danbooru. All trademarks remain the property of their
respective owners. e621 and Danbooru tag names are used for interoperability and are the property
of their respective communities.
License
This fine-tune is released under the Apache License 2.0.
The base model prithivMLmods/Qwen3-VL-8B-Instruct-c_abliterated-v3
derives from Qwen3-VL-8B-Instruct, which is also Apache-2.0. Please consult the base model
repository for its authoritative license terms and any additional restrictions introduced by the
abliteration process.
The e621 and Danbooru tag vocabularies are community-maintained resources. This repository
does not redistribute either vocabulary — it is used as a labelling standard only.
Acknowledgements
- Qwen Team for Qwen3-VL-8B-Instruct.
- prithivMLmods for the
c_abliterated-v3checkpoint this fine-tune builds on. hearmeneighfor thee621-rising-v3-smalldataset.- The e621 and Danbooru communities for the tag standards that make this model's output
interoperable with the wider ecosystem. - The llama.cpp project for the GGUF format and multimodal runtime.