Wan video LoRA training

Preview ecosystem

Wan video training is currently marked Preview in the orchestrator. The endpoint accepts requests and whatif=true cost previews work, but actual training runs may not be available on every worker fleet. Reach out via Civitai Discord before integrating against production traffic.

Train a WAN video LoRA on a small set of source video clips using AI Toolkit. Output is a video LoRA usable in WAN text-to-video and image-to-video generation.

modelVariant	Wan family	Default price
2.1	Wan 2.1 (14B)	3000 Buzz
2.2	Wan 2.2 (14B-A14B)	3000 Buzz

Video training is the slowest training mode on the platform — a 2000-step run on a 4-clip dataset takes many minutes. Submit with wait=0. The request fields, training data, continuing a run and live progress are shared by every ecosystem — see the LoRA training overview.

The request shape

{
  "$type": "training",
  "input": {
    "engine":       "ai-toolkit",
    "ecosystem":    "wan",
    "modelVariant": "2.1"        // 2.1 | 2.2
  }
}

Prerequisites

• A training-data zip of images, video clips, or a mix — addressed by an AIR (accepted sources), with an accurate count of items. Keep clips to a few seconds at a similar resolution.

Wan 2.1 / 2.2

Both variants share the same input shape and per-step cost; pick the one that matches your inference target. The example below uses 2.1; swap modelVariant to "2.2" for Wan 2.2 training (no other change required).

POST https://orchestration.civitai.com/v2/consumer/workflows?wait=0
Authorization: Bearer <your-token>
Content-Type: application/json

{
  "tags": ["training", "video"],
  "steps": [{
    "$type": "training",
    "priority": "normal",
    "retries": 2,
    "input": {
      "engine": "ai-toolkit",
      "ecosystem": "wan",
      "modelVariant": "2.1",
      "steps": 2000,
      "resolution": 512,
      "lr": 0.0002,
      "trainTextEncoder": false,
      "lrScheduler": "constant",
      "optimizerType": "adamw8bit",
      "networkDim": 32,
      "networkAlpha": 32,
      "trainingData": {
        "type": "zip",
        "sourceUrl": "urn:air:other:other:huggingface:datasets/Civitai/orchestration-samples@e86d025874700507615ae8ef74937c319ea41dfe/sample-training-dataset.zip",
        "count": 15
      },
      "samples": {
        "prompts": ["a video of TOK", "TOK moving in a garden"]
      }
    }
  }]
}

POST/v2/consumer/workflows

Set your Civitai API token via the Token button in the navbar to enable Try It.

Request body — edit to customize (e.g. swap the image URL or prompt)

{ "tags": [ "training", "video" ], "steps": [ { "$type": "training", "priority": "normal", "retries": 2, "input": { "engine": "ai-toolkit", "ecosystem": "wan", "modelVariant": "2.1", "steps": 2000, "resolution": 512, "lr": 0.0002, "trainTextEncoder": false, "lrScheduler": "constant", "optimizerType": "adamw8bit", "networkDim": 32, "networkAlpha": 32, "flipAugmentation": false, "trainingData": { "type": "zip", "sourceUrl": "urn:air:other:other:huggingface:datasets/Civitai/orchestration-samples@e86d025874700507615ae8ef74937c319ea41dfe/sample-training-dataset.zip", "count": 15 }, "samples": { "prompts": [ "a video of TOK", "TOK moving in a garden" ] } } } ] }

Valid JSON

Wan parameters

Fields and defaults not listed here are as in the overview's common parameters.

Field	Wan default	Notes
modelVariant	✅ required	2.1 or 2.2.
epochs	round(steps / 200)	Each adds 200 Buzz — Wan's per-epoch preview samples are videos and expensive to compute, so keep the epoch count modest.
batchSize	1	Fixed at 1 for this ecosystem.
continueFrom	(none)	A Wan LoRA (urn:air:wan:lora:...) to resume from.
lr	0.0001	0.0002 is a typical override for video; see example.
trainTextEncoder	false	Leave off — Wan training does not benefit from text-encoder updates.
triggerWord	(none)	Some video ecosystems reject triggerWord — leave it empty if you see schema rejections.
trainingData	✅	Images, video clips, or a mix; the example uses an image dataset.
samples.prompts[]	[]	Preview videos rendered at each saved checkpoint with the trained LoRA.

Reading the result

The standard training output: each saved checkpoint yields a video LoRA .safetensors blob plus any sample .mp4 files. The trained LoRA is usable in WAN video generation by referencing it in the loras field.

Runtime

Wall time, default settings on a 4-clip dataset:

Variant	Per 100 steps	Typical full run
2.1	~1–3 min	20–60 min for 2000 steps
2.2	~1–3 min	20–60 min for 2000 steps

Cost

The training cost formula with Wan rates:

price = steps × costPerStep + epochs × 200       (rounded)
costPerStep = 0.5   (2.1 and 2.2)
floor: never less than 80% of the default-configuration price

Wan's per-epoch surcharge is 200 Buzz (not 10 like image ecosystems) because each epoch's preview samples are videos and expensive to compute — so most of Wan's cost is in the epoch count, not the step count. With both steps and epochs omitted the run is 2000 steps / 10 epochs (the example at the top of the page sends steps: 2000, which also resolves to 10 epochs) → 2000 × 0.5 + 10 × 200 = 1000 + 2000 = 3000 Buzz. The floor is 80% of the default price (2400 Buzz).

Lowering epochs saves the most. Sample-prompt rendering itself uses Wan video-generation rates and is billed separately. Preview the exact charge with whatif=true (Payments (Buzz)).

Configuration	Buzz (training only)
default (steps: 2000, epochs: 10)	3000 + samples
steps: 2000, epochs: 20	5000 + samples (each extra checkpoint adds 200)
steps: 1000, epochs: 5	2400 + samples (floor)

Troubleshooting

steps out of range, an unresolvable sourceUrl, a count mismatch and a rejected dataset are covered in the overview's troubleshooting.

Symptom	Likely cause	Fix
400 with "modelVariant required"	Missing modelVariant	Set to "2.1" or "2.2".
Step starts then fails immediately	Wan training not yet available for your request	Contact Civitai support.
Step failed with VRAM-related error	Resolution × clip length too high for the worker	Lower resolution (e.g. to 512), shorten clips to ≤ 3 seconds.
Trained LoRA produces static / no motion	Too few steps, too few / too short clips	Raise steps; ensure clips show the motion you want learned.

Related

• LoRA training overview — shared parameters, training data, live progress and cost formula
• LTX2 video LoRA training — Lightricks LTX video LoRA training (also video, less expensive previews on LTX2.3)
• WAN video generation — use a trained LoRA in WAN inference
• Flux 2 Klein LoRA training — image-side counterpart
• Results & webhooks
• SubmitWorkflow · GetWorkflow
• training OpenAPI spec