flammenai/FlammenSDXL1-GGUF

FlammenSDXL1

Anime-style SDXL merge — the production image-generation checkpoint behind flammen.ai (PWA dating-style chat with AI characters). Published as GGUF for use with stable-diffusion.cpp.

Validated against several alternative candidates (NikuMixXL5, NikuMix2.5 SD1.5, NikuXLv0.1) on a same-seed / same-prompt grid for prompt fidelity and skin-tone adherence; FlammenSDXL1 won on character variety and the failure modes most relevant to a character-generation product.

Files

File	Size	Notes
FlammenSDXL1-f16.gguf	~7 GB	Unquantized. Highest quality; lets you re-quantize into any sd-cpp format.
FlammenSDXL1-q8_0.gguf	~4 GB	Production weight. Near-lossless. Recommended for most users.
FlammenSDXL1-q4_1.gguf	~3 GB	Consumer-grade quantization. Runs on 6–8 GB cards (3060, 4060). Some quality loss vs q8.

All quants keep the VAE at f16 — SDXL VAEs degrade noticeably under aggressive quantization.

Inference (stable-diffusion.cpp)

sd-cli \
  -m FlammenSDXL1-q8_0.gguf \
  -p "anime style, masterpiece, best quality, …, <your tags>" \
  -n "<see negative prompt below>" \
  -W 768 -H 768 \
  --steps 22 --cfg-scale 7 \
  --sampling-method euler_a \
  -o out.png

/sdapi/v1/txt2img (when running sd-server) accepts the same parameters in the standard A1111 JSON shape.

Recommended generation parameters

These are the production settings used to render every Flame avatar on flammen.ai.

• Resolution: 768 × 768 (the current OOM ceiling on a 4060 Ti; higher works if you have the VRAM)
• Steps: 22
• CFG scale: 7
• Sampler: Euler a
• VAE: the f16 VAE is baked in; no separate VAE needed

Perf on a single 4060 Ti (16 GB), q8_0, 768², 22 steps: ~9.6 s per image. Model uses ~4 GB VRAM + ~4.3 GB compute buffer for VAE decode — pin to a GPU with at least 9 GB free.

Prompting

SDXL anime checkpoints carry strong default-archetype bias (femme, light-skinned, East-Asian) that overrides descriptor tags unless those tags are protected by counter-tags in the negative prompt. The settings below are the hard-won prod recipe — start here and tune.

Positive prompt template

anime style, masterpiece, best quality, high detail,
solo, looking at viewer, simple background, upper body,
<your image tags>

Base negative prompt

lowres, bad anatomy, bad hands, text, error, missing fingers,
extra digit, fewer digits, cropped, worst quality, low quality,
jpeg artifacts, signature, watermark, username, blurry,
photo, photorealistic, 3d, realistic, multiple people

Conditional negative additions (append to base)

• If generating a male character, append 1girl, female, woman, breasts, cleavage, large breasts (the model gravitates female by default).
• If generating a female character, append 1boy, male, man, beard, mustache, facial hair.
• For dark-skinned characters (tags like dark_skin, very_dark_skin), append pale_skin, light_skin, white_skin, fair_skin (and tan for very dark) — the model fights non-default skin tones aggressively without this.
• For pale-skinned characters (when the model defaults to a tan), append dark_skin, dark-skinned_male, dark-skinned_female, tanned_skin.

License

creativeml-openrail-m (the standard SDXL base license). You are free to use the model for inference, derivatives, and commercial purposes subject to the use-based restrictions in the license — read the full text before deploying.

Provenance

Internal merge by @nbeerbower of the NikuMix line. Quantized to q8_0 GGUF for stable-diffusion.cpp deployment. Currently serving production traffic on a Witchgrid-managed sd-server cluster.