NUILab Agentic AI
Docs / Models / What's installed
Models
What's installed
Every model on disk, what it is for, its size, and its licence.
Two tiers. Tier A fits alongside the resident services (~8.7 GB budget). Tier B needs
comfy-up to evict them first (~15.3 GB budget).
Licences are as published on each model’s page at the time of download. Check the licence yourself before shipping anything commercially — this table is a convenience, not legal advice, and several otherwise-excellent models in this space are non-commercial.
Tier A — the working set
FLUX.2 klein 4B — the primary generator
| File | Size | Licence |
|---|---|---|
diffusion_models/flux-2-klein-4b-fp8.safetensors | 4.07 GB | Apache 2.0 |
diffusion_models/flux-2-klein-4b-nvfp4.safetensors | 2.46 GB | Apache 2.0 |
text_encoders/qwen_3_4b_fp4_flux2.safetensors | 3.85 GB | Apache 2.0 |
vae/flux2-vae.safetensors | 0.34 GB | Apache 2.0 |
A 4-billion-parameter model that does generation and multi-reference editing in one. Apache 2.0, so there is no licence problem shipping what it makes. The whole stack peaks around 5 GB, which is why it is the default rather than SDXL.
Two builds are installed on purpose. nvfp4 is a 4-bit format that Blackwell executes
natively — this card is the specific hardware that makes it worth having. fp8 is the safer,
more widely supported build. Start with fp8; try nvfp4 when you want speed or headroom.
Z-Image Turbo — the ControlNet lane
| File | Size | Licence |
|---|---|---|
diffusion_models/z_image_turbo_nvfp4.safetensors | 4.51 GB | Apache 2.0 |
text_encoders/qwen_3_4b_fp4_mixed.safetensors | 3.48 GB | Apache 2.0 |
vae/ae.safetensors | 0.34 GB | Apache 2.0 |
model_patches/…Controlnet-Union-2.1-lite-2602-8steps.safetensors | 2.02 GB | Apache 2.0 |
A 6B model with the best union ControlNet in the open ecosystem. It is here because klein has none — see Choosing a model .
Warning
models/model_patches/, not models/controlnet/. Putting it
in the obvious directory means ComfyUI silently will not list it.SDXL — the long tail
| File | Size | Licence |
|---|---|---|
checkpoints/sd_xl_base_1.0.safetensors | ~6.9 GB | CreativeML OpenRAIL++-M |
vae/sdxl_vae.safetensors | ~0.33 GB | (see repo) |
controlnet/diffusion_pytorch_model_promax.safetensors | 2.51 GB | Apache 2.0 |
loras/pixel-art-xl.safetensors | 170 MB | CreativeML OpenRAIL-M |
Kept for the enormous back catalogue of community LoRAs and for ControlNet-Union ProMax, which covers more control modes than anything newer.
Pixel-art and sprite LoRAs
| File | Size | Base | Licence |
|---|---|---|---|
loras/pytorch_lora_weights.comfyui.safetensors | 325 MB | klein-4B | Apache 2.0 |
loras/flux-spritesheet-lora.safetensors | 72.5 MB | klein-4B | Apache 2.0 |
loras/pixel_4walk_small_flux2_klein_base_4b_v1.safetensors | 92.4 MB | klein-4B base | Apache 2.0 |
loras/pixel-art-xl.safetensors | 170 MB | SDXL | CreativeML OpenRAIL-M |
The first produces sprites on a transparent background, which removes an entire cut-out step. The second generates a 2×2 turnaround — four camera angles of one design. The third is a 4-direction walk cycle sheet at 32×32.
Background removal and upscaling
| File | Size | Licence |
|---|---|---|
background_removal/birefnet.safetensors | 424 MB | (upstream MIT) |
upscale_models/RealESRGAN_x4plus_anime_6B.pth | ~18 MB | BSD-3-Clause |
upscale_models/4x-PixelPerfectV4.pth | 63.9 MB | WTFPL |
BiRefNet is the matting model for cutting sprites out when the transparent-background LoRA has not
already done it. ⚠ It lives at background_removal/birefnet.safetensors inside its repo, not at
the repo root — fetching the root path returns 404.
Deliberately the only learned upscaler installed. See Pixel-perfect output for why an upscaler is usually the wrong tool for pixel art, and for the licence trap that eliminated the popular alternatives.
Tier B — needs eviction
| File | Size | Licence |
|---|---|---|
diffusion_models/z_image_turbo_bf16.safetensors | 12.31 GB | Apache 2.0 |
model_patches/…Controlnet-Union-2.1-2602-8steps.safetensors | 6.71 GB | Apache 2.0 |
Full-precision Z-Image and the full-size ControlNet. Better output than the quantised tier-A pair,
and they do not fit unless you run comfy-up.
Deliberately not installed
| Model | Why not |
|---|---|
| FLUX.2-dev (32B) | 64 GB, and non-commercial licence |
| Qwen-Image-2512 | 20.4 GB fp8 — does not fit |
| Qwen-Image-Layered | 20.5 GB — does not fit, though RGBA layer output is genuinely attractive |
| Qwen-Image-Edit-2511 | Best open identity-preserving editor, but 20B; only a heavily-quantised build fits and it needs a further ~9 GB text encoder |
| Ideogram 4.0 | Fits at ~12 GB, but the licence is non-commercial |
| Pixel Art Diffusion XL | Creator restricts redistribution and selling outputs |
Fetching or re-fetching
~/scripts/fetch-art-models.sh a # tier A only
~/scripts/fetch-art-models.sh b # tier B only
~/scripts/fetch-art-models.sh all
Idempotent — it compares each local file against the published size, so complete files are skipped
and interrupted ones resume. A clean re-run reports 0 downloaded, 19 skipped, 0 failed.
Warning
content-length header of the HEAD
response. Do not use curl’s %{size_download} with -I: a HEAD transfers no body, so it always
reports 0, every file looks incomplete, and the script re-runs the entire set. It is a quiet bug
— curl -C - no-ops on an already-complete file, so nothing breaks and nothing is corrupted; the
script simply lies about having downloaded 48 GB. This one shipped here and was caught by
re-running and noticing the skip count was wrong.Source: content/models/catalog.md · maintained in the nuilab-aigaming repository.