NUILab Agentic AI
Docs / Models / Choosing a model
Models
Choosing a model
Which of the installed models to reach for, and how the landscape shifted in 2026.
The short answer
| You want | Use |
|---|---|
| A sprite, icon, portrait, or item | klein-4B + a pixel LoRA |
| The same character from four angles | klein-4B + the spritesheet LoRA |
| A specific pose or composition you control | Z-Image Turbo + its union ControlNet |
| A style only an old community LoRA can do | SDXL + that LoRA |
| The best quality you can get on this card | comfy-up, then Z-Image bf16 (tier B) |
Why not “just use SDXL”
Because that answer expired. SDXL was the obvious default for years on the strength of its LoRA ecosystem, and that ecosystem still matters — which is why it is installed. But a 4B Apache-2.0 model that generates and edits, in 5 GB, beats it on the axes that matter here: memory, licence clarity, and multi-reference consistency.
The rule that survives: the model with the deepest LoRA library wins for style, and the model with the best conditioning wins for control. Those are no longer the same model, which is why this box has three.
The trap: klein has no ControlNet
This is the single most important compatibility fact in the set.
FLUX.2 klein has no union ControlNet. It relies on reference-latent conditioning instead.
So if you need explicit pose or composition control — a character standing in a specific stance, art matched to a layout — you cannot use the primary model. Your options are SDXL + ControlNet-Union ProMax or Z-Image Turbo + Union 2.1. Both are installed for exactly this reason.
What changed in 2026, briefly
Worth knowing because a lot of advice written before this year now points at the wrong things:
- FLUX.2 arrived, and the small
kleinvariants are Apache 2.0 — an unusually permissive licence for a model this capable. The 32Bdevvariant is not; it is non-commercial. - Qwen went closed. Qwen-Image 2.0 and 3.0 shipped without weights. Anything newer than the late-2025 releases is not available to run locally, no matter what a blog post implies.
- Purpose-built sprite LoRAs became a category. Transparent-background sprite output and spritesheet-editing LoRAs did not meaningfully exist before this year; both are installed.
- Quantisation got hardware-specific. On Blackwell, prefer NVFP4 builds over GGUF — the card executes 4-bit natively rather than emulating it.
- FLUX 3 was announced but is API-only. Open weights are “coming soon”. Do not plan around it.
What is still genuinely hard
Single-frame art is largely solved at this scale. Two things are not, and no amount of model selection fixes them:
- Consistent animation frames. Frame-to-frame coherence remains weak. Budget for hand cleanup.
- Truly seamless tilesets. Output that tiles perfectly without visible seams still usually needs a human pass.
Plan the pipeline around that rather than being disappointed by it.
Source: content/models/choosing.md · maintained in the nuilab-aigaming repository.