← Docs Home · Model Catalog · HTTP API · Troubleshooting

Wan 2.1 T2V 1.3B LoRA Inference with Diffusers (AI Toolkit-trained)

API model id: wan21_1b
URL slug: wan21-1b

This page documents the reference Diffusers inference pipeline for wan21_1b (Wan 2.1 T2V 1.3B). It is designed for running LoRAs trained with ostris/ai-toolkit while minimizing training preview vs inference mismatch. If you are trying to reproduce AI Toolkit sample previews, treat the code linked below as the source of truth (scheduler wiring, resolution snapping, LoRA application, and conditioning).

Run in the cloud (optional): If you want to reproduce the examples on this page in a pinned runtime without local CUDA/driver setup (and reduce preview‑vs‑inference drift), run it via RunComfy’s Cloud AI Toolkit (Train + Inference). 👉 You can open it here: Cloud AI Toolkit (Train + Inference)

Quick facts

Field	Value
Pipeline	`src/pipelines/wan21.py`
Base checkpoint	`Wan-AI/Wan2.1-T2V-1.3B-Diffusers`
Defaults	`sample_steps=25`, `guidance_scale=4.0`, `seed=42`
Resolution snapping	Floors width/height to a multiple of 16
Control image	No
Video	Yes (`num_frames=41`, `fps=16` by default)
LoRA scale behavior	Dynamic via adapters (`set_adapters`). Scale is set per request via `loras[].network_multiplier`.
Needs AI Toolkit	No

Reference implementation (source of truth)

Pipeline implementation: src/pipelines/wan21.py
Shared behaviors (snapping, seeding, LoRA base logic): src/pipelines/base.py
Request schema (parameter names): src/schemas/request.py
Model ids (enum): src/schemas/models.py
Pipeline registry (model → class): src/pipelines/__init__.py
Executor (prompt processing + caching): src/tasks/executor.py

Minimal API request

{
  "model": "wan21_1b",
  "trigger_word": "sks",
  "prompts": [
    {
      "prompt": "[trigger] a photo of a person",
      "width": 1024,
      "height": 1024,
      "seed": 42,
      "sample_steps": 25,
      "guidance_scale": 4.0,
      "neg": "",
      "num_frames": 41,
      "fps": 16
    }
  ],
  "loras": [
    {
      "path": "my_lora_job/my_lora.safetensors",
      "network_multiplier": 1.0
    }
  ]
}

Pipeline behavior that matters

Same API surface as Wan 2.1 14B, but with a smaller base checkpoint (lower VRAM requirement).
Uses diffusers.WanPipeline for text-to-video.

Preview-matching notes (training preview vs inference mismatch)

Match num_frames and fps when comparing outputs.
Width/height are floored to a multiple of 16.

What to compare when debugging mismatch

Confirm the effective width/height after snapping (the server floors to the divisor).
Match sample_steps and the scheduler family (FlowMatch / UniPC / DDPM differences matter).
Match guidance_scale semantics (some pipelines map it to a different internal parameter).
Match loras[].network_multiplier and whether LoRA scale is dynamic vs fused.
If video: match num_frames and fps (and any frame-count constraints).