SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsmodel-routing

model-routing

Agents & models

View repository

Updated 10-29 16:15

About this skill

Renoise Model Routing

Select by task, not by habit. Then prompt for the selected model instead of sending every model the same generic brief.

Hard Boundary

Live model capabilities are authoritative for availability, defaults, input roles, combinations, limits, duration, ratio, resolution, and audio controls. This Skill contains only researched routing preferences and prompting styles.

  1. Preserve a model explicitly named by the user, or one you presented as selected in the plan or estimate. If it cannot do the job, explain why and the replacement you propose, and switch only after the user agrees.
  2. Filter candidates by the live capabilities required by the task.
  3. Route only to the newest live generation of a model family; older generations remain available only when the user explicitly names one.
  4. Choose the best specialist below.
  5. Use the live isDefault model only when no specialist clearly fits or candidates tie.
  6. Inspect the selected model's live guidance before writing the final prompt.
  7. Let the selected model's profile override generic prompt-craft advice when prompt structure, density, or reference wording conflicts.
  8. Never pass a role or parameter merely because this guide mentions a public model capability; the Renoise deployment may expose a narrower contract.

Live availability alone is not a routing recommendation. Omit older generations, fully dominated models, and models without a clear task advantage; an explicitly named model still follows rule 1.

References and Edits

  • Keep the authoritative identity or product references across iterative edits: send them in every related task alongside the latest composition or edit source, state each reference's job, and never treat a transformed intermediate as the replacement. When a reference limit forces a choice, keep the original anchor.
  • For an edit or continuation, keep the source's aspect ratio when the model supports it; analyze the source when the ratio is unknown, and follow the model's live guidance for source-following ratios.
  • Before generating with references, check every requirement the live capabilities state (format, duration, size, count) against what analysis or metadata shows. A source that breaks one would be rejected: convert or trim it with code first.

Routing Questions

Classify the request before selecting:

  • Kind: image, video, or audio.
  • Operation: create, edit, continue, interpolate, animate a reference, or upscale.
  • Priority: final quality, strict instruction following, aesthetics, identity consistency, speed, or cost.
  • Structure: exact text/layout, one cinematic shot, multi-shot narrative, dialogue performance, music, or a complete audio scene.
  • References: none, one source image, several identity/style references, source video, endpoint frames, or voice/audio references.

Do not call a model “best” without naming the task it is best for.

Image Routing

Decision Order

Task Prefer Why
Exact text, UI, diagrams, ads, packaging, compositing, identity-sensitive image-to-image work, or peak final quality gpt-image-2.5-sunburst Route a generic GPT Image 2.5 request here; use gpt-image-2.5-flare when explicitly requested. Image references guide transformations, but neither variant exposes mask inpainting.
General high-quality photorealistic or commercial image live image default Start with the deployment's current default when no specialist better fits; do not infer local-edit or typography guarantees from the model family.
Fast, inexpensive drafts and simple variants grok-image Lowest-cost live image tier for quick iteration; move the chosen direction to a specialist for the final.
Extreme aspect ratios (1:4 to 8:1), Google-family work, several references, multilingual localization, or conversational iteration nano-banana-2 The Google-family model in the live lineup, balancing quality, latency, text, reference reasoning, and broad formats.
Aesthetic exploration, stylized art direction, editorial mood, or concept art mj-v8.2 Current Midjourney generation and strong aesthetic prior.
Seedream output resolution or image-reference count beyond what the live Pro contract accepts seedream-5-0-lite Choose Lite when its advertised limits fit and Pro's do not; otherwise use the live default for general Seedream work.
xAI image request grok-image for lower-cost iteration; grok-image-quality when higher quality justifies the added cost Both support direct natural-language generation and editing through the live Renoise contract.

Image Prompting Styles

GPT Image 2.5 — production brief

For complex production briefs, use labeled, ordered instructions:

USE: [ad / packaging / UI / infographic / edit]
SCENE/BACKGROUND: ...
SUBJECT: ...
COMPOSITION: ...
LIGHTING/MATERIALS: ...
EXACT TEXT: "..." with placement, hierarchy, font character, and color
PRESERVE: identity, geometry, layout, logo, and all unmentioned elements
CHANGE ONLY: ...
AVOID: ...
  • Route a generic GPT Image 2.5 request to Sunburst; use Flare when explicitly requested. Neither live contract supports mask inpainting.
  • Put literal in-image text in quotes and specify hierarchy and placement.
  • For image-to-image transformations, say change only X; preserve everything else.
  • Identify every reference by its job: source, identity, product, layout, or style.
  • Prefer one-change iterations over rewriting the whole brief; do not promise mask inpainting.

Nano Banana 2 — conversational creative direction

[Subject] + [action] + [location] + [composition] + [style]
Camera/viewpoint: ...
Lighting/materials: ...
Exact text: "..."
Keep unchanged: ...
  • Start with a clear operation verb and use positive, concrete natural language.
  • For references, state each source's job, the relationship among them, and the new scenario.
  • For edits, separate the requested change from what remains unchanged; repeat preservation requirements on later turns.
  • Quote exact visible copy and specify its hierarchy, placement, visual character, and localization target.
  • Relate multiple references explicitly and iterate conversationally.

Seedream 5 — design brief

Put the deliverable format first, then subject, layout, light, exact text, and invariants:

[DELIVERABLE FORMAT and ratio/use case]
Subject and setting: ...
Layout and spatial relationships: ...
Lighting, materials, and palette: ...
Exact visible copy: "..."
Keep exactly: ...
  • Pro: specify dense layout, multilingual copy, materials, local edit regions, and everything the edit preserves.
  • Lite: natural language works well; make causal/spatial relationships explicit for complex transformations.
  • Assign each reference a clear role and avoid vague pronouns in multi-reference edits.

Midjourney — aesthetic visual phrase

Use a concise visual description, not a requirements document:

[subject], [medium], [environment], [lighting], [palette], [mood], [composition]
  • Prefer concrete visual nouns and adjectives over keyword spam.
  • Name the lighting and composition that matter instead of relying on “cinematic.”
  • Use Midjourney for look development rather than text-heavy layouts or surgical edits.

Grok Imagine Image — concise scene direction

[subject doing action] in [setting], [camera/framing], [lighting], [material/style]
  • Use standard for lower-cost iteration and Quality when the higher-quality tier is worth the added cost.
  • For edits, attach the source image or images and describe the requested change directly.
  • Keep the prompt concrete enough that the main subject and action remain dominant.
Video Routing

Decision Order

Task Prefer Why
General multimodal video, continuous 16–30 second sequences, rich mixed references, source-video edits, or forward/backward extension seedance-2.5-byteplus Newest Seedance generation and the long-form, reference-heavy specialist with precise edit/extension workflows.
Budget multimodal video: brief shots up to 30 seconds and 1080p with a few image/video/audio references, text-to-video, or first(/last)-frame generation, when Seedance 2.5's long-form, edit, or extension strengths are not needed wan3.0-video About a third of Seedance 2.5's per-second price with broad reference support; it cannot edit, extend, or continue a source video, and frame and reference modes do not mix.
Same as Wan 3.0 when turnaround matters more than cost wan3.0-video-prime Identical capabilities, faster on longer and higher-resolution jobs, and about 1.5× Wan 3.0's per-second price; prefer plain Wan 3.0 otherwise.
Short 720p generation, text rendering, or source-video edit within its narrow live contract gemini-omni-flash Strong short-form instruction following, multi-shot generation, and conversational editing.
H3-family 2K delivery, high-fidelity endpoint control, or focused multimodal references hailuo-h3 Full-quality tier with image/video/audio references; choose it when Max's live contract cannot meet the task.
Fast H3-family text-to-video, first/last-frame, or reference-to-video with image/video/audio sources hailuo-h3-max Faster multimodal tier when its live quality and resolution fit; suitable for source-video variants without discarding the source.
Fastest, lowest-cost H3-family text-to-video or first/last-frame generation without generic references h3-max-turbo Turbo supports text and endpoint frames, not reference image/video/audio inputs; use H3 Max or H3 for references.
A single shot delivered at 2K or 4K from text or first(/last) frames, with native sound and no other references flux-3-video The only live model with 4K output (2K/4K upscaled from the same generation), 5–20 seconds, and cheaper per second than Seedance 2.5 at 1080p; it takes no reference images, video, or audio.
xAI one-image animation with native sound grok-video-1.5 Newest Grok video generation; the current Renoise contract requires exactly one input image.
Digital human: one person image talking or singing to a given voice track (talking head, presenter monologue, dubbed avatar, lip-synced song) omnihuman-2.0 Lip-syncs one portrait to one driving audio and runs as long as the audio (up to 300 seconds), so the exact voice and long monologues are controllable; it does not stage product handling or camera action the way generative video does.
Upscale an existing video without changing its content upscale-video-topaz-starlight-2.5 Dedicated restoration/upscaling model; use a generation or editing model when objects, motion, timing, or style should change.

Tier and version notes

  • Within the H3 family, use full H3 when its higher-quality output or 2K matters; choose H3 Max when speed/cost matters and its live image/video/audio reference roles, combinations, and output quality fit. Turbo is the text/first/last-frame budget tier, not a generic reference mode. A source video needs reference_video; first_frame alone is not equivalent.
  • Within FLUX 3, use FLUX 3 Video for 2K/4K or 9:21; choose flux-3-max when the shot stays at 720p/1080p and the user asks for FLUX's higher-quality tier, at a higher per-second cost. Independent rankings of FLUX 3 disagree, so do not prefer it over Seedance 2.5 on quality alone.
  • Grok's public upstream capabilities may move faster than Renoise's contract. Route from live roles and limits rather than assuming an upstream feature is connected.
  • Model preference leaderboards are task-, resolution-, and audio-filter-specific; use them as volatile evidence, not a single aggregate ranking.

Video Prompting Styles

Seedance 2.5 — multimodal director brief

Assign each reference an explicit job, then write shots:

References:
- @material:<ID>: subject identity
- @material:<ID>: location/style
- @material:<ID>: motion/camera/voice

Shot 1: framing, subject action, camera move, environment, sound/dialogue.
Shot 2: ...
Continuity: traits and objects that must remain unchanged.
Constraints: no subtitles/logo/watermark unless requested.
  • Use a few stable identity features rather than an exhaustive biography.
  • Give each beat a clear camera behavior; avoid simultaneous conflicting moves.
  • Describe body part, speed, force, and physical consequence for actions.
  • Externalize emotion through visible behavior and direct dialogue/audio explicitly.
  • Begin with the intended result, map every reference to its purpose, then use non-overlapping timestamp ranges or numbered shots with continuity notes.
  • For a source-video edit, name the source, the intended change, its time range when relevant, and what remains unchanged.
  • For extension, state forward or backward direction and describe the visual, motion, and audio continuity across the source boundary.

Wan 3.0 — single-shot brief

[subject and action], [setting], [one camera move], [lighting/style].
References: @material:<ID> identity / @material:<ID> motion / @material:<ID> sound.
Audio: dialogue, effects, or ambience — or no audio.
  • Keep each job to one brief shot or a short causal sequence; route multi-beat narratives, edits, and extensions to Seedance 2.5.
  • Write @material:<ID> mentions; the live contract rewrites them into Wan's numbered image/video/audio labels.
  • Choose frame mode (first frame, optionally with a last frame) or reference mode, never both; a last frame alone is rejected.
  • Never ask Wan to edit, extend, continue, remove, or replace content in an existing video.

Gemini Omni Flash — conversational generation/editing

  • For creation: write a direct natural-language brief with subject, action, camera, setting, visible text, and desired audio.
  • Gemini readily creates multi-shot clips; say “single continuous shot,” “single unbroken scene,” or “no scene cuts” when continuity is required.
  • Use natural time phrases or [0–3s] blocks for important beats.
  • For editing: identify the source, say change only X, and list what remains unchanged.
  • Make one edit per turn when possible; use follow-up refinement rather than replacing the entire prompt.

MiniMax H3 family — mode-specific audiovisual plan

Choose one mode that the selected model exposes. Full H3 and H3 Max can use generic reference mode when their live roles allow the specific image/video/audio mix; H3 Max Turbo supports text and first/last-frame modes, not generic references. Frame and reference roles are separate modes — do not combine them unless the selected model's live guidance explicitly permits it:

  • Text-to-video: state the subject, action over time, camera movement, setting, lighting, and sound intent directly.

  • First frame: describe only the motion, camera path, action development, and sound after the supplied opening state.

  • Last frame: describe the plausible path that converges on the supplied ending.

  • First + last: describe the continuous transition between states; avoid re-describing the two stills. Prefer one coherent shot unless a cut is essential.

  • Reference mode (H3 / H3 Max when live): assign each image, video, and audio an explicit identity, motion, style, voice, action, or sound job. Attach the authorized source video directly when its motion/timing matters, and check the selected model's live duration and role-combination limits.

For complex reference work use:

References: each @material:<ID> image/video/audio and its job.
Task summary: target result and what each source contributes.
Preserve: identity, motion, voice, style, or content that carries over.
[Shot 1] Visual action, camera motion, dialogue, and diegetic sound.
[Shot 2 at time] Cut and continuation.
Overall soundscape: ambience, Foley, and non-verbal sound.
Non-diegetic music: instrumentation, tempo, dynamics — or none.

Specify camera type + amplitude + speed only when they matter. Keep dialogue exact, label speakers consistently, and separate in-scene sound from background score.

Grok Imagine Video — animate the change

For image-to-video, do not re-describe the static source:

[subject motion/change], [one camera move], [atmosphere/lighting change].
Sound: specific dialogue/SFX/ambience; no music if unwanted.
  • Use one primary action or a short causal sequence and front-load the important motion.
  • Attach exactly one input image and describe the change from that starting frame.
  • Give the attached image a clear visual or motion job.
  • Request dialogue, effects, ambience, or music explicitly when native audio matters.

OmniHuman 2.0 — the audio drives the performance

The voice track is the script: write and generate the audio first (usually seed-audio-1.0 from the exact lines, or the user's authorized recording), check it is speech or vocals of 4–300 seconds, then pass it with one portrait.

  • Use a single, front-facing or three-quarter portrait with the face and mouth unobstructed; the framing of the image is the framing of the video.
  • The prompt is optional; when used, describe expression, gestures, and subtle camera behavior, never new dialogue.
  • Do not pass duration, ratio, or resolution: length follows the audio and output is 720p.
  • Pure instrumental audio is rejected; for a song, pass the vocal track.

Topaz Starlight — no creative prompt

Upscaling is not regeneration. Supply only the source video and settings required by the live capability; do not add a creative prompt or unrelated media inputs. If the user wants content changed, route to a video editing model instead.

Audio Routing
Task Prefer Why
Song, instrumental, score cue, jingle, or music bed — anything where music is the deliverable lyria-pro Composes a complete structured song (verses, chorus, bridge, optional vocals and lyrics) up to about 3 minutes; replaces Lyria 3 Clip.
A vocal song sung from exact lyrics (the user's, or ones you write from their theme), optionally with a female or male lead mureka-song-9 Dedicated sung-lyrics model: lyrics go in their own field with section tags and the style stays in the prompt; one complete song of about 2.5–3 minutes.
Dialogue, expressive narration, podcast, dubbing/re-voicing, or a complete sound scene seed-audio-1.0 Unified speech, dialogue, effects, and ambience generation; live default audio model.
One standalone sound effect: foley, impact, whoosh, UI sound, ambience loop, or stinger eleven-sfx-v2 Exact length from 1 to 30 seconds and billed per second; use Seed Audio when the effects belong to a scene with speech.

Lyria composes music and stays the choice for instrumentals, cues, and beds; Mureka sings given lyrics; Seed Audio handles speech-led and complete sound scenes; ElevenLabs SFX makes single effects to length.

Lyria 3 Pro — composer brief

Genre/style: ...
Mood: ...
Instrumentation: ...
Tempo/rhythm: ... BPM
Vocals: instrumental OR voice type, delivery, and language
Lyrics/theme: exact lyrics or subject
Target length: ~N seconds (approximate)
Structure: [Intro] → [Verse] → [Chorus] → [Bridge] → [Outro]
Production: mix character, era, space, dynamics
  • Explicitly say instrumental when vocals are unwanted.
  • Name instruments, tempo, vocal range/texture, backing vocals, and language.
  • There is no duration parameter: state the target length in the prompt and treat it as approximate; match lyric volume to that length.
  • Label sections with [Intro], [Verse], [Chorus], [Bridge], and [Outro]; put exact lyrics under the section they belong to.
  • Use timestamps only for meaningful structural changes; musical alignment follows bars rather than sample-accurate timing.
  • Describe genre, era, instrumentation, and production traits instead of requesting a named artist imitation.
  • Use an optional guide image only when the live material roles expose it.

Mureka Song 9 — lyrics plus a style line

lyrics:
[Verse]
...
[Chorus]
...
prompt: genre, mood, tempo, instruments, vocal character
  • Put the full lyrics in the lyrics field with section tags; never put lyrics in the prompt.
  • Size the lyrics to a 2.5–3 minute song; the length is not adjustable.
  • Choose the lead's gender only when it matters; otherwise let the model pick.

Seed Audio 1.0 — sound-scene script

Setting/acoustics: ...
Continuous sound bed: ...
Speaker A (age, accent, timbre, emotion, pace): "Exact line."
Action-tied SFX: ...
Speaker B (...): "Exact line."
Music cue and ending: ...
  • For full-scene work, direct it as a scene; for speech-only work, specify voice identity, performance, pacing, and acoustics.
  • Label speakers and exact lines; match prompt language to dialogue language.
  • Use clean authorized voice references, or an authorized character image for inferred voice, according to the mutually exclusive live reference modes.
  • Use concrete Foley and ambience; onomatopoeia can clarify transient sounds.
  • Use precise timing for dialogue; describe SFX, ambience, and music timing approximately unless live guidance says otherwise.

ElevenLabs Sound Effects — one sound, described physically

Describe the source, action, material and texture, and the space (for example "heavy oak door slams shut in a stone hallway, short echo"). Set the duration to the effect's real length, and ask for one effect per task; layer several effects in the edit.

Research Basis

Reviewed 2026-10-13 against live Renoise model contracts. Rankings are directional and decay quickly; live capabilities and task-specific tests override them.

Primary guidance: