# Skill: Pick the right open-source image/video model (2026) for the task

> Choose the correct open-weights model for a generation task instead of defaulting to one model or chasing hype; apply whenever a user asks "which model should I use" for local image or video generation in ComfyUI.

Source: sodigi-learn - comfyui/karta-modeley-2026 - https://sodigi.io/learn/comfyui/karta-modeley-2026

## When to use

- A user wants to generate images or video locally (ComfyUI / open weights) and is unsure which model fits.
- Output quality is poor because the chosen model does not match the task (e.g. garbled text, weak photorealism).
- Hardware is limited and a heavy model fails or is too slow.
- Someone asks to use a closed API model (Ideogram, Midjourney, Wan 2.5) "in ComfyUI".

## Core rules

- Task first, model second. Define the task in one word (photo / text / video / art) before picking anything.
- Text inside the image -> Qwen-Image (and Qwen-Image-Edit for editing existing text). Best open option for legible text.
- Top quality / brand-consistent series -> FLUX.2 (dev for quality + multi-reference consistency; klein for speed/low VRAM).
- Video -> Wan 2.1 / 2.2 (open, ComfyUI-native; text-to-video and image-to-video).
- Huge detailed scenes / 1000-word prompts -> HunyuanImage 3.0 (80B MoE, needs strong GPU).
- Weak GPU or need speed -> Z-Image (6B), FLUX.2 klein, Stable Diffusion 3.5, or a quantized (GGUF) build of any model.
- Open-weights = downloadable, runs locally, free. Ideogram and Wan 2.5 are closed APIs - cannot run locally in ComfyUI.
- Download ONE model per task. Do not hoard checkpoints.

## Procedure

1. Name the task: photo, text-in-image, video, or art.
2. Check available VRAM on the GPU.
3. Map task -> model: text=Qwen-Image, video=Wan, quality=FLUX.2, giant scenes=HunyuanImage.
4. If VRAM is low, pick the light variant (Z-Image, FLUX.2 klein, SD 3.5) or a GGUF/quantized build of the chosen model.
5. Confirm the model is open-weights (not a closed API) before promising local use.
6. Download only that one model; verify it loads; then build the workflow.

## Ready-to-use prompt

```
I need to generate <describe output> locally in ComfyUI. My GPU has <N> GB VRAM.
Recommend ONE open-weights model for this, following these rules:
- text inside the image -> Qwen-Image; editing existing image/text -> Qwen-Image-Edit
- top quality / consistent series -> FLUX.2 (dev or klein by my VRAM)
- video (text-to-video or image-to-video) -> Wan 2.1/2.2
- very large detailed scene -> HunyuanImage 3.0 (only if VRAM is high)
- low VRAM / need speed -> Z-Image, FLUX.2 klein, SD 3.5, or a GGUF build
Tell me the model, why it fits, the lighter variant if my VRAM is tight, and confirm it is open-weights (not a closed API like Ideogram or Wan 2.5).
```

## Pitfalls

- Recommending the newest/biggest model regardless of the user's GPU (it will not load).
- Confusing open-weights with closed APIs: Ideogram, Midjourney, Wan 2.5 cannot run locally.
- Using a text-weak model for posters/signage instead of Qwen-Image.
- Suggesting users download many models "just in case" - wastes disk; two or three working ones suffice.
- Treating prompt wording as a fix for a wrong model choice.