Skip to content

recommend_training

catalog

Pick a Scenario training base architecture: image or video LoRA family + variant, or voice clone, from a user training intent.

Not in the default tool list

This tool lives in the tool catalog. Discover it with scenario_tools_search (or enumerate everything with scenario_tools_list) and run it via scenario_tool_execute_read. To list every tool directly, connect with ?toolsets=full.

When to use

Reach for this when the user asks "which model should I train" / "what LoRA base fits my dataset" / "voice clone for X". Since it's catalog-only, find it with scenario_tools_search and run it via scenario_tool_execute_read — it returns the exact type string to pass to model_create. For picking models that generate outputs, use recommend instead.

Decision tree

The picker filters by dataset shape, scores by style + subject, then picks a variant inside the winning family by priority.

Image families

FamilyKindBest forStatus
Flux 2standardNew projects training a single-image LoRA — characters, products, environments, or distinctive styles.stable
Flux 2 EditeditWhen your dataset is image pairs with text instructions describing the change.stable
Qwen ImagestandardHigh-volume use cases where budget matters and you want strong prompt adherence.stable
Qwen EditeditEdit recipes that should keep the surrounding context intact.stable
Z-ImagestandardStylized work — character art, illustration, concept art — or anywhere the highest training fidelity matters.stable
Flux 1 DevstandardExisting projects with Flux 1 LoRAs already in production. New projects should use Flux 2.legacy
Flux KontexteditExisting edit recipes already using Flux Kontext. Consider Flux 2 Edit or Qwen Edit for new projects.legacy

Video families

Pass modality: "video" to train a video LoRA. The dataset shape picks the family and priority the variant; style and subject do not apply. A clip dataset never gets an images-only variant, and with video, single_images means images only.

Video training families
FamilyDataset shapeVariantsBest for
MiniMax H3images_or_video_clips
minimax-h3-lora
minimax-h3-int8-lora (images only)
Video LoRAs that learn motion and style from your images or video clips.
MiniMax H3 Referenceclips_with_references
minimax-h3-ref-lora
When each clip comes with the references it was made from (a character sheet, a game screenshot) and you will give the model similar references at generation time.

Subject to the MiniMax H3 license. Its outputs may not be used to train other AI models.

Voice options

  • Instant Voice Cloning (IVC) (stable) — Quick voice clone from a short audio sample. Ready in seconds.
  • Professional Voice Clone (soon) — High-fidelity clone trained on longer, cleaner audio — better expressiveness and realism.

Required inputs

At minimum, pass prompt describing the training goal — the LLM extracts everything else. Hosts that already know the modality + dataset shape can pass them as structured params and skip LLM 1.

Output shape

{
  family: "Flux 2",
  recommended: {
    type: "flux.2-dev-lora",
    label: "Dev (32B)",
    why: "..."
  },
  alternatives: [
    { type, when_better, tradeoff }
  ],
  dataset_requirements: {
    shape: "single_images",
    recommended_size: "5-15 images",
    max_size: 50
  },
  family_notes?: string,
  resolution_notes: string,
  _degraded?: "llm_unavailable" | "llm_timeout"
}

Worked examples

Image — character LoRA

Prompt: "I have 30 photos of my dog. Train a model."

  • family: Flux 2
  • recommended: flux.2-dev-lora
  • dataset: 5–15 single images, max 50

Voice — quick clone

Prompt: "15-second voice memo, want a quick clone."

  • family: Instant Voice Cloning (IVC)
  • recommended: voice-clone-ivc
  • dataset: short audio sample, <30s

Reference

Public training docs: docs.scenario.com/get-started/training/training-models