executing-plans
executing-plans
Use when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available
Explore skills
Productivity · SKILL.md
higgsfield-models
higgsfield-modelsUse when the user asks which model to use, wants to compare models, or needs guidance on selecting between Kling, Wan (incl. Wan 3.0), Seedance (incl. 2.5), Veo 3, Minimax Hailuo / MiniMax H3, FLUX 3 Video, Gemini Omni Flash, Soul, Nano Banana, GPT Image 2.5, or other Higgsfield engines — including which lane edits existing footage (video_edit, Kling 3.0 Om…
Independent GitHub snapshot · Automated classification · Original documentation governs
Skill documentation
GitHubname: higgsfield-models description: > Use when the user asks which model to use, wants to compare models, or needs guidance on selecting between Kling, Wan (incl. Wan 3.0), Seedance (incl. 2.5), Veo 3, Minimax Hailuo / MiniMax H3, FLUX 3 Video, Gemini Omni Flash, Soul, Nano Banana, GPT Image 2.5, or other Higgsfield engines — including which lane edits existing footage (video_edit, Kling 3.0 Omni Edit, FLUX 3 Video Edit, Genjutsu, Ad Multiplier) and which models make one clip longer than 15s. user-invocable: true metadata: references: - MODELS-DEEP-REFERENCE.md tags: [higgsfield, models, Kling, Sora, Wan, Seedance, Veo, Soul, NanoBanana, GPT-Image-2.5, FLUX-3, Genjutsu, edit-lanes, long-take] version: 3.3.3 updated: 2026-09-26 parent: higgsfield
Higgsfield Model Selection Guide
Choosing the right model is the single biggest factor in output quality after the prompt.
This file handles most selection questions. For deep per-model documentation (prompting
specifics, parameters, edge cases, API details) → read MODELS-DEEP-REFERENCE.md.
The canonical comparison tables, the Edit-Lane Chooser, the Long-Take Chooser, and
the Motion Transfer table live in ../../model-guide.md — this file agrees with them.
Star ratings are written only from real generations. A — in a star column means not
yet field-rated; platform facts for those rows are [OFFICIAL — platform, snapshot 2026-09-26].
Quick Decision Flowchart
Fast lookup — for detailed comparisons see the full tables below.
| Need | Recommended Model | Tier |
|---|---|---|
| Top-tier cinematic video + audio | Kling 3.0 | Premium |
| Epic scale / spectacle | Seedance 2.0 or Minimax Hailuo 2.3 (Sora 2 is retired — OpenAI shut its API down 2026-09-24) | Premium |
| Nature / landscapes | Veo 3.1 | Premium |
| Artistic / stylized video | Wan 2.6 | Mid |
| One clip longer than 15s | Seedance 2.5 · Wan 3.0 / Prime · FLUX 3 Video (to 20s) → model-guide.md § Long-Take Chooser | — |
| Edit / extend / multiply existing footage | → model-guide.md § Edit-Lane Chooser | — |
| Motion transfer from a reference video | Kling 3.0 Motion Control or Genjutsu → model-guide.md § Motion Transfer | — |
| Fast video iteration | Seedance 2.0 Fast / Mini | Mid |
| VFX / fluid motion | Minimax Hailuo 2.3 | Mid |
| Budget-friendly video | Kling 3.0 Turbo — the catalog's budget Kling (Kling 2.5 Turbo / Higgsfield DoP Lite: not in the API catalog, 2026-09-26 — verify in the live UI) | Mid |
| Fashion / aesthetic images | Soul 2.0 | Free |
| Photorealistic sharp images | Nano Banana Pro | Low |
| AI actor generation | Soul Cast | Low |
| Native 4K images | Nano Banana Pro (resolution to 4k) — Kling Image 3.0 is not in the API catalog, 2026-09-26 — verify in the live UI | — (NB Pro's listed price is for 1K; verify the 4K price) |
| Transparent-background image | GPT Image 2.5 (background: transparent) | — |
| Photo style transformation | Photodump (29 presets) | Low |
Pricing tiers: Free (Soul 2.0; DoP Lite — not in the API catalog, verify in the UI) · Low (0.1–2 credits) · Mid (2–10 credits) · Premium (10+ credits). See the Credit Cost Reference below for exact per-model costs.
Video Models — Comparison
| Model | Realism | Character | Motion | Style | Duration | Audio | Best for |
|---|---|---|---|---|---|---|---|
| Kling 3.0 | ★★★★★ | ★★★★★ | ★★★★★ | ★★★★☆ | 3–15s | ✅ | Cinematic, long, audio, multi-shot |
| Kling 3.0 Omni | ★★★★★ | ★★★★★ | ★★★★★ | ★★★★☆ | 3–15s | ✅ | Video clone, storyboard control — not in the API catalog, 2026-09-26 — verify in the live UI (only its edit model kling_video_edit is). Storyboards: Kling 3.0 |
| Kling 3.0 Omni Edit | ★★★★★ | ★★★★★ | — | ★★★★☆ | 3–10s in (UI doctrine) | ✅ | Edit footage at 3.0 quality — kling_video_edit, mode std / pro / 4k, source video + optional image refs |
| Kling O1 Video (legacy) | ★★★★★ | ★★★★★ | ★★★★☆ | ★★★☆☆ | 5–10s | ❌ | Multi-ref (7), start/end frame — not in the API catalog, 2026-09-26 — verify in the live UI. Start/end frame: Kling 3.0; many refs: Seedance 2.0 / 2.5 |
| Kling O1 Video Edit (legacy) | ★★★★☆ | ★★★★★ | — | ★★★★★ | 3–10s | ❌ | Relight, restyle, swap, remove — UI-only, not in the API catalog |
| Kling 3.0 Motion Control | ★★★★★ | ★★★★☆ | ★★★★★ | ★★★☆☆ | 3–30s | Optional | Motion transfer from reference video |
| Genjutsu — motion transfer | — | — | — | — | — | — | hf_mult_motion_control: motion from a reference video onto subjects in reference images; 480p–1080p. Not yet field-rated |
| Genjutsu — replace object | — | — | — | — | — | — | hf_mult_replace_object: replace objects in a source video from reference images; 480p–1080p. Not yet field-rated |
| Kling 2.6 (legacy) | ★★★★★ | ★★★★★ | ★★★★☆ | ★★★☆☆ | 5/10s | ✅ | Character drama, realism; native audio via sound toggle (default on) |
| Kling 2.5 Turbo | ★★★★☆ | ★★★★☆ | ★★★★☆ | ★★★☆☆ | 5–10s | ❌ | Fast Kling iteration — not in the API catalog, 2026-09-26 — verify in the live UI. Catalog: Kling 3.0 Turbo |
| Sora 2 (retired) | ★★★★☆ | ★★★☆☆ | ★★★★★ | ★★★★☆ | 4–12s | ✅ | Retired — do not recommend. OpenAI shut the Sora 2 API down on 2026-09-24; Higgsfield UI availability is unconfirmed. Scale / physics → Seedance 2.0 or Minimax Hailuo 2.3 (../../model-guide.md). Was: epic scale, physics, action — UI-only |
| Wan 3.0 | — | — | — | — | 2–30s or −1 smart (billed as 10s) | ✅ | T2V, first/last frame, multimodal reference (image/video/audio), enable_thinking; frames and references never combined. Not yet field-rated |
| Wan 3.0 Prime | — | — | — | — | 2–30s or −1 smart (billed as 10s) | ✅ | Same parameter surface as Wan 3.0; how Prime differs is not stated. Not yet field-rated |
| Wan 2.7 | ★★★★★ | ★★★★☆ | ★★★★★ | ★★★★★ | 2–15s | ✅ | 60fps, T2V/I2V/R2V/edit, first+last frame |
| Wan 2.6 | ★★★★☆ | ★★★☆☆ | ★★★★☆ | ★★★★★ | 5/10/15s | ❌ | Artistic, stylized, improved physics |
| Wan 2.5 | ★★★★☆ | ★★★☆☆ | ★★★★☆ | ★★★★★ | 5–10s | ✅ | Native audio, artistic, fantasy — not in the API catalog, 2026-09-26 — verify in the live UI. Catalog: Wan 2.6 (stylized) / Wan 2.7 (native audio) |
| Seedance 2.5 | — | — | — | — | 4–30s | ✅ | Omni-reference (≤30 images incl. start/end, ≤50 items), video_edit, video_extension; 480p / 720p / 1080p; start/end frames only in omni_reference; no 4K, no genre. Not yet field-rated |
| Ad Multiplier | — | — | — | — | 4–30s | ✅ | "Powered by Seedance 2.5" — many edited variants of one 4–30s ad via Higgsfield's ad-multiplier workflow. Not yet field-rated |
| Seedance 2.0 | ★★★★★ | ★★★★★ | ★★★★★ | ★★★★☆ | 4–15s | ✅ | 12-asset multimodal, complex motion, native 4K (mode=std), genre |
| Seedance 2.0 Fast | ★★★★☆ | ★★★★☆ | ★★★★★ | ★★★★☆ | 4–15s | ✅ | mode=fast of 2.0 — no 1080p/4K |
| Seedance 2.0 Mini | ★★★★☆ | ★★★★☆ | ★★★★☆ | ★★★★☆ | 4–15s | ✅ | Budget tier, full reference roles + genre; no 1080p/4K |
| Seedance 1.5 Pro | ★★★★☆ | ★★★★☆ | ★★★★☆ | ★★★★☆ | 4/8/12s | ✅ | Best lip-sync, multilingual audio |
| Seedance Pro (legacy UI label) | ★★★☆☆ | ★★★☆☆ | ★★★☆☆ | ★★★☆☆ | 10s | ❌ | Not in the API catalog — use Seedance 2.0 Fast / Mini |
| Veo 3.1 | ★★★★★ | ★★★★☆ | ★★★★☆ | ★★★★☆ | 4/6/8s | ✅ | Ref images, first/last frame, 4K (Google API — the catalog's veo3_1 shows only start_image; verify) |
| Veo 3.1 Lite | ★★★★☆ | ★★★★☆ | ★★★★☆ | ★★★★☆ | 4/6/8s | ✅ | Budget 3.1 quality, 1080p, start + end frame, volume |
| Veo 3 | ★★★★☆ | ★★★☆☆ | ★★★★☆ | ★★★☆☆ | 4–8s | ✅ | Nature, environment, stable model |
| Gemini Omni Flash | ★★★★☆ | ★★★★☆ | ★★★☆☆ | ★★★☆☆ | 4–10s | ✅ | Reference-driven video (image + video refs), native audio, 720p |
| Gemini Omni Flash 1.1 | — | — | — | — | 3–10s | ✅ | Required mode (t2v / i2v / reference / edit — source ≤30s); start/end frames; 360p–4K. Not yet field-rated |
| Grok Video | ★★★★☆ | ★★★☆☆ | ★★★★☆ | ★★★★☆ | 1–15s | ✅ | Animate images, social clips (grok_video, named Grok Imagine through the 2026-06-22 snapshot; the catalog exposes no source-video input — editing unverified on Higgsfield) |
| Minimax Hailuo 2.3 | ★★★★★ | ★★★★☆ | ★★★★★ | ★★★★☆ | 6/10s | ❌ | VFX, fluid motion, anime, physics |
| Minimax Hailuo 02 | ★★★★☆ | ★★★☆☆ | ★★★★★ | ★★★☆☆ | 6/10s | ❌ | Dance, sports, fluid motion — no catalog variant is named 02 (minimax_hailuo: minimax, minimax-fast, minimax-2.3, minimax-2.3-fast); verify in the live UI |
| MiniMax H3 | — | — | — | — | 4–15s | — | Keyframes or image/video/audio refs, 2K, batch_size 1–4. Not yet field-rated |
| MiniMax H3 Max | — | — | — | — | 5–15s | — | "Fast" variant, 480p / 768p, same roles, batch_size 1–4. Not yet field-rated |
| Happy Horse Video | — | — | — | — | 3–15s | — | T2V + single start frame, 720p / 1080p. Not yet field-rated |
| FLUX 3 Video | — | — | — | — | 5–20s | ✅ | T2V, multi-frame I2V, continuation; native 2:1. Not yet field-rated |
| FLUX 3 Video Edit | — | — | — | — | first 15s of source | — | Text edit of one video, 1 credit per processed second. Not yet field-rated |
| Higgsfield DoP (Lite/Standard/Turbo) | ★★★☆☆ | ★★★☆☆ | ★★★★☆ | ★★★☆☆ | 3–5s | ❌ | I2V specialist, 50+ presets, optical physics — not in the API catalog, 2026-09-26 — verify in the live UI |
Decision Flowchart
Is this image or video?
├── IMAGE
│ ├── Person / portrait? → Soul 2.0
│ ├── Cinematic keyframe for I2V pipeline? → Soul Cinema (`soul_cinematic`) ·
│ │ Soul Cinema Preview (no catalog model by that name — verify in the live UI)
│ ├── Native 4K? → Nano Banana Pro · image series / storyboarding? → Kling
│ │ Image 3.0 (not in the API catalog, 2026-09-26 — verify in the live UI)
│ ├── Maximum sharpness / 4K? → Nano Banana Pro
│ ├── Fast pro-quality / text rendering? → Nano Banana 2
│ ├── Reference consistency or dense text? → Seedream 4.5
│ ├── Complex layout / multi-panel? → Seedream 5.0 Lite
│ ├── Text/logo in image? → GPT Image 2 (or 2.5 — not yet field-rated; see higgsfield-gpt-image-2)
│ ├── Transparent-background cut-out? → GPT Image 2.5
│ ├── Extend / crop the canvas per side? → FLUX.2 Pro Outpaint
│ ├── Masked inpaint? → Nano Banana 2 / 2 Lite (`mask` + `is_inpaint`)
│ └── Edit an existing image? → Flux Kontext
│
└── VIDEO
├── EXISTING footage to edit, extend, or multiply?
│ └── → model-guide.md § Edit-Lane Chooser (one table, every lane)
│
├── One clip longer than 15s?
│ └── → model-guide.md § Long-Take Chooser
│ (Seedance 2.5 · Wan 3.0 / Prime · FLUX 3 Video · Cinema Studio 4.0)
│
├── Is a human character the focus?
│ ├── Need audio, up to 15s, multi-shot → Kling 3.0
│ ├── Need to clone from reference video → Kling 3.0 Omni (not in the API
│ │ catalog, 2026-09-26 — verify in the live UI; no catalog model on record)
│ ├── Best lip-sync + multilingual → Seedance 1.5 Pro
│ ├── Legacy-tier great character (audio togglable via `sound`) → Kling 2.6
│ └── Fast iteration → Kling 3.0 Turbo (Kling 2.5 Turbo: not in the API catalog)
│
├── Need motion transfer from reference video?
│ └── → Kling 3.0 Motion Control or Genjutsu (model-guide.md § Motion Transfer)
│
├── Animate a still image with cinematic camera?
│ └── → Kling 3.0 (`start_image`); Higgsfield DoP (Lite/Standard/Turbo) is
│ not in the API catalog, 2026-09-26 — verify in the live UI
│
├── Is the environment/phenomenon the hero?
│ ├── Nature, documentary, stable → Veo 3
│ ├── Need ref image consistency → Veo 3.1 (verify refs in the UI)
│ ├── Budget Veo 3.1 quality / volume → Veo 3.1 Lite
│ ├── 60fps, first+last frame, ref images → Wan 2.7
│ └── Artistic, painterly, fantasy → Wan 2.6 (Wan 2.5: not in the API catalog)
│
├── Is it action/spectacle?
│ ├── Epic scale, crowds, physics → Seedance 2.0 (Sora 2 is retired)
│ ├── VFX, anime, fluid motion → Minimax Hailuo 2.3
│ └── Dance, sports, budget motion → Minimax Hailuo 2.3 (02: no catalog
│ variant by that name — verify in the live UI)
│
├── Need maximum reference control?
│ ├── Up to 30 images / 50 items (images + video + audio) → Seedance 2.5
│ ├── Up to 12 assets with 4K or `genre` → Seedance 2.0
│ ├── Up to 7 image refs → Kling O1 Video (not in the API catalog — verify;
│ │ the two Seedance rows above take more)
│ └── Image + video refs with native audio → Gemini Omni Flash / 1.1
│
└── Speed/cost priority?
├── Fastest Kling → Kling 3.0 Turbo (Kling 2.5 Turbo: not in the API catalog)
├── Seedance drafts → Seedance 2.0 Fast / Mini
├── Up to 4 takes per call → MiniMax H3 / H3 Max (`batch_size`)
└── Fastest Veo → Veo 3.1 Fast or Veo 3 Fast (`variant: veo-3-1-fast` /
`veo-3-fast` — the catalog defaults)
Image Models — Quick Selection
| Need | Model | Credits |
|---|---|---|
| Fashion / cultural portrait | Soul 2.0 | Free |
| Cinematic keyframe for I2V | Soul Cinema (soul_cinematic) · Soul Cinema Preview (no catalog model by that name, 2026-09-26 — verify in the live UI) | — · Low |
| Consistent character identity (16:9) | Soul Cast | budget 10–500 |
| Environment / location plate | Soul Location | — |
| Cheapest generation | Z-Image | 0.15 |
| Low-cost portrait | Soul 2.0 · Higgsfield Soul (not in the API catalog, 2026-09-26 — verify in the live UI) | Free · 0.5 |
| Low-cost 2K square | Kling O1 Image (kling_omni_image) | 0.5 |
| Native 4K / image series | Nano Banana Pro (4K) · Kling Image 3.0 (series; not in the API catalog, 2026-09-26 — verify in the live UI) | 2 at 1K (4K: verify) · — |
| 4K + advanced editing | Nano Banana 2 (edits, to 4k) · Kling Image 3.0 Omni (not in the API catalog, 2026-09-26 — verify in the live UI) | 1.5 at 1K (4K: verify) · — |
| Fast versatile 2K | Seedream 5.0 Lite | 1 |
| Fast generation + instruction editing up to 2K | Seedream 5.0 Flash (not yet field-rated) | — |
| 4K versatile | Seedream 4.5 | 1 |
| Sketch-to-image (Draw) | Nano Banana | 1 |
| Artistic / stylized | Seedream 5.0 Pro (stylized-2D) · Wan 2.2 (not in the API catalog, 2026-09-26 — verify in the live UI) | — · 1 |
| Blend multiple references | Nano Banana Pro (14 refs) · Multi Reference (not in the API catalog, 2026-09-26 — verify in the live UI) | 2 · 1.5 |
| Fast pro-quality + text rendering | Nano Banana 2 | 1.5 |
Budget NB2 (1k only, thinking MINIMAL/HIGH) | Nano Banana 2 Lite | — |
Transparent background · quality to max · 15 aspect ratios | GPT Image 2.5 (not yet field-rated) | — |
| Complex prompts / text in image | GPT Image 2 · GPT Image 1.5 (left the API catalog after 2026-06-22 — verify in the live UI) | — · 2 |
| Reference-based editing + best text rendering | OpenAI Hazel | — |
| Max fidelity / Thinking mode / 14 refs | Nano Banana Pro | 2 |
| xAI generation + editing | Grok Image / Grok Image 2.0 (not yet field-rated) | — |
| Image editing / inpainting | Flux Kontext | varies |
| Extend or crop the canvas per side | FLUX.2 Pro Outpaint (not yet field-rated) | — |
| Photo style transformation (29 cartoon/illustration presets) | Photodump | Low |
Full image model specs + UI controls → ../../image-models.md
Full Photodump preset library (29 named styles) → ../../photodump-presets.md
Budget Tiers
Image models — by credit cost:
- Free / near-free: Soul 2.0 (5K gens) · Z-Image (0.15) · Face Swap (2 free)
- Budget (0.5–1): Kling O1 Image · Seedream family · Nano Banana · Higgsfield Soul, Wan 2.2, Reve (these three: not in the API catalog, 2026-09-26 — verify in the live UI)
- Mid (1.5–2): Nano Banana 2 · FLUX.2 Pro · Flux Kontext Max (UI tier — the API's
flux_kontextis now named Flux Kontext) · NB Pro · Character Swap · Multi Reference, GPT Image (not in the API catalog, 2026-09-26 — verify in the live UI) - Premium (5–6): FLUX.2 Flex · FLUX.2 Max
General pricing tiers (video + image, approximate):
- Free: Soul 2.0 · DoP Lite (limited; not in the API catalog, 2026-09-26 — verify in the live UI)
- Low: 0.1–2 credits per generation
- Mid: 2–10 credits per generation
- Premium: 10+ credits per generation
For exact per-model video costs see the Credit Cost Reference in ../../model-guide.md.
Unique Feature Matrix
Catalog-backed rows cite [OFFICIAL — platform, snapshot 2026-09-26] (media roles and params in
../../specs/MODEL-SPECS.md / IMAGE-MODEL-SPECS.md); other entries are earlier doctrine.
| Feature | Available on |
|---|---|
| Native audio (dialogue, SFX, ambient) | Kling 3.0/2.6 · Seedance 2.5/2.0/2.0 Mini/1.5 Pro · Ad Multiplier · Wan 3.0/Prime · FLUX 3 Video · Gemini Omni Flash / 1.1 · Veo 3/3.1/3.1 Lite · Wan 2.7 · Grok Video · Cinema Studio 3.0 / 4.0 · (Kling 3.0 Omni, Wan 2.5 — not in the API catalog, 2026-09-26 — verify in the live UI) |
| Soul ID character slot | Soul 2.0 · GPT Image, Higgsfield Soul (not in the API catalog, 2026-09-26 — verify in the live UI) |
| @ Elements syntax | Seedream 4.5/5.0 Lite · Nano Banana Pro · Cinema Studio |
| Draw (sketch-to-image) | Nano Banana · Nano Banana Pro |
| Video editing (existing footage) | Seedance 2.5 video_edit · Cinema Studio 4.0 video_edit · Kling 3.0 Omni Edit · FLUX 3 Video Edit · Gemini Omni Flash 1.1 edit · Genjutsu replace-object · Ad Multiplier (many variants) · Kling O1 Video Edit (UI-only legacy) → ../../model-guide.md § Edit-Lane Chooser |
| Multi-image reference blend | Seedance 2.5 (≤30 images) · Nano Banana Pro (14 refs) · Multi Reference, Kling O1 Video (7 refs) — not in the API catalog, 2026-09-26 — verify in the live UI |
| Start/end frame control | Seedance 2.5 (omni_reference only) / 2.0 / 2.0 Mini / 1.5 Pro · Kling 3.0 · Wan 3.0/Prime (not with references) · Wan 2.7 · FLUX 3 Video · Gemini Omni Flash 1.1 · MiniMax H3 / H3 Max · Minimax Hailuo · Veo 3.1 Lite · Cinema Studio 3.0 / 4.0 · Marketing Studio · Ad Multiplier · Kling O1 Video (legacy — not in the API catalog) |
| Video extension | Seedance 2.5 / Ad Multiplier / Cinema Studio 4.0 video_extension (forward / backward) · FLUX 3 Video (continuation) · Veo 3.1 (Google API chain to 148s — the catalog's veo3_1 takes no video input; verify) |
| One clip longer than 15s | Seedance 2.5 (30s) · Wan 3.0/Prime (30s) · FLUX 3 Video (20s) → ../../model-guide.md § Long-Take Chooser |
| Performance cloning from video | Kling 3.0 Omni — not in the API catalog, 2026-09-26 — verify in the live UI; no catalog model is on record for likeness + voice cloning |
| Motion transfer from a reference video | Kling 3.0 Motion Control (3–30s reference) · Genjutsu hf_mult_motion_control |
| Soul Cast AI actors | Cinema Studio 2.5 · standalone soul_cast image model (16:9, budget 10–500) |
| Soul Cast AI actors (General 2K / Character 4K / Location 4K) | Cinema Studio 3.0 (Business/Team) |
| Built-in color grading | Cinema Studio 2.5 (full grading suite) · Cinema Studio 3.5 (Color Palette axis in Style Settings — 8 named palettes) |
| Native dual-channel stereo audio | Cinema Studio 3.0 (Business/Team) · Kling 3.0 · Seedance 2.0/1.5 Pro · Veo 3/3.1 · Wan 2.7 · (Kling 3.0 Omni, Wan 2.5 — not in the API catalog) |
| Soul HEX color matching | Soul 2.0 · Cinema Studio 2.5 · Soul Cinema Preview (no catalog model by that name — verify in the live UI) |
| Native 4K image series | Kling Image 3.0 — not in the API catalog, 2026-09-26 — verify in the live UI (native 4K alone: Nano Banana Pro) |
| Style presets + Color Transfer | Soul 2.0 (the CLI forbids style_id + image references in one Soul 2.0 call — see ../../image-models.md § Soul 2.0) |
| Transparent-background image output | GPT Image 2.5 background: transparent (the CLI also lists background on GPT Image 2 — see ../../image-models.md § GPT Image 2) |
| Masked inpaint | Nano Banana 2 / 2 Lite (mask + is_inpaint) · GPT Image 2 (CLI only). Seedream 5.0 Pro has is_inpaint (edit the reference) but no mask role |
| Google Search grounding | Nano Banana Pro |
| Negative prompts | Veo 3/3.1 · Wan 3.0 (Alibaba docs — URLs in MODELS-DEEP-REFERENCE.md § Wan 3.0: an in-prompt "Negative prompt list" section, not a separate parameter) |
| Batch of up to 4 takes per call | MiniMax H3 / H3 Max (batch_size) |
| Smart auto-camera planning | Cinema Studio 3.0 (Business/Team) |
Key Model Notes
Kling 3.0 vs 2.6: 3.0 is the current top Kling model — longer clips (15s vs 10s), native audio, multi-shot AI direction, physics engine, 4K HDR, stylized output engine. 2.6 is now legacy — use 3.0 for all new work unless cost is the primary constraint.
Kling V3 vs O3: Use V3 for prompt-driven cinematic work (text-to-video, image-to-video).
Use O3 when you have reference media (video or image+audio) to anchor character identity —
O3's reference-based consistency is its defining advantage. O3 (Kling 3.0 Omni) is not in the API
catalog as of the 2026-09-26 snapshot — only its edit model, kling_video_edit, is; verify it in the
live UI before recommending.
Kling 3.0 Motion Control: Upload a 3–30s reference clip to transfer full-body motion,
hand gestures, facial expressions. Image Orientation for camera/talking head; Video
Orientation for complex motions (dancing, action, full-body movement). Genjutsu
(hf_mult_motion_control) is the other motion-transfer lane — see model-guide.md § Motion Transfer.
Seedance 2.0: Rule of 12 (up to 12 assets per generation). Real person face uploads blocked — use synthetic character references. Best practices for Seedance 2.0 prompting are integrated into the sub-skills (see higgsfield-prompt, higgsfield-camera, higgsfield-motion).
Seedance 2.5 (2026-09-26 surface): 4–30s, 480p / 720p / 1080p, four modes. start_image /
end_image are accepted only in omni_reference; t2v takes no media. 4K and the genre
param remain on the 2.0 family (2.0; genre also on 2.0 Mini). Dialect: higgsfield-seedance-2-5.
Wan 3.0 / 3.0 Prime: 2–30s (or -1 smart duration, billed as 10s — offer it only when the
user asks the model to pick the length), 480p–1080p, native audio, enable_thinking. A call
carries either frames or references, never both. Vendor prompting dialect:
MODELS-DEEP-REFERENCE.md § Wan 3.0.
Veo 3.1 vs 3.1 Lite vs 3: 3.1 adds reference images (up to 3), first/last frame, video extension, 4K
at the Google API; the Higgsfield catalog's veo3_1 exposes only a start_image role, so verify
those in the UI. 3.1 Lite is budget-priced 3.1 quality at 1080p — supports T2V and I2V with start +
end frames, costs less than half of 3.1 Fast. 3 is stable and proven. Use 3.1 for subject
consistency, 3.1 Lite for volume, 3 for pure environment/nature.
Wan 2.7: Major upgrade — native 60fps (vs 24fps in 2.6), up to 15s duration, first+last frame anchoring, up to 5 reference images, 4-model suite (T2V/I2V/R2V/video edit), Flow-Matching architecture. 40% better physics consistency over 2.6.
Wan 2.5: First Wan version with native audio — joint text/audio/video generation. Supports audio-driven video (upload audio to drive visuals). 1080p, 5–10s. Not in the API catalog as of the 2026-09-26 snapshot; may be UI-only — verify in the live UI before recommending (catalog Wan with native audio: Wan 2.7).
Minimax Hailuo 2.3 vs 02: 2.3 is a major upgrade — improved physics, anime/illustration styles, facial micro-expressions, better prompt adherence. Fast variant now at 1080p (02 Fast was 512p). 02 was kept for budget motion work, but no catalog variant is named 02 (2026-09-26: minimax_hailuo offers minimax, minimax-fast, minimax-2.3, minimax-2.3-fast) — verify in the live UI before recommending it.
Grok Imagine: Aurora architecture (autoregressive, not diffusion) — excels at text/logo
rendering and multi-image compositing. On Higgsfield: Grok Image (grok_image) and
Grok Image 2.0 (grok_image_2_0) are live image models; Grok Video takes only a start image.
Gemini Omni Flash / 1.1: Google's reference-driven video models. 1.0 accepts image and video
references, native audio, 4–10s at 720p (16:9 or 9:16). 1.1 adds a required mode
(text-to-video / image-to-video / reference-to-video / edit), start/end frames, 3–10s, and
360p–4K; its edit mode uses the source duration capped at 30s. 1.1 is not yet field-rated.
For deep documentation on any specific model → read MODELS-DEEP-REFERENCE.md
Cinema Studio 3.0 (Business/Team Plan)
Cinema Studio 3.0 is a separate generation engine available on Business and Team plans. Version toggle in the upper-right corner of the Cinema Studio UI switches between 2.5 and 3.0.
| Feature | Cinema Studio 2.5 | Cinema Studio 3.0 (Business/Team) | Cinema Studio 3.5 |
|---|---|---|---|
| Video Resolution | Up to 1080p | Up to 720p (may increase) | 480p / 720p / 1080p (three-tier) |
| Image Resolution | Up to 4K | Up to 4K (Character/Location) · Up to 2K (General) | 1.5K / 2K (Soul Cinema, default image model) · 1K / 2K / 4K (Cinematic Cameras image model) |
| Max Duration | 12s | 15s | 15s |
| Aspect Ratios | 6 options | 7 options (+ 21:9 ultrawide) | Video: 7 options (Auto, 16:9, 9:16, 4:3, 3:4, 1:1, 21:9) · Image: 8 options (1:1, 3:4, 2:3, 9:16, 3:2, 4:3, 16:9, 21:9) |
| Audio | On/Off | On/Off (native dual-channel stereo) | On/Off (generated alongside video) |
| Shot Control | Manual multi-shot | Smart (auto) + Custom multi-shot | Video: 3-pill main UI (Genre / Style / Camera) · Image: Cinematic models picker (Soul Cinema default + Cinematic Characters / Locations / Cameras) — see higgsfield-cinema |
| Generation Cost | Varies | 48 credits | Varies — see Higgsfield plan documentation |
For full Cinema Studio 3.0 documentation → see
higgsfield-cinema
Cinema Studio 3.5 sits alongside 2.5 and 3.0 in the model selector — all three coexist on the platform, version is user-selected, and there is no auto-routing between them. 3.5 reframes the surface: the main UI collapses creative control into three pills (Genre / Style / Camera), each defaulting to Auto with manual override available. Optical physics is restored via a four-axis Camera Settings panel (Camera Body / Lens / Focal Length / Aperture, with 75mm added as a new focal length vs 2.5's 8/14/35/50mm set — vocabulary differs from 2.5; do not mix). The Style Settings panel exposes three preset axes (Color Palette / Lighting / Camera Moveset Style) plus a free-form Manual Style mode for natural-language style direction. An AI director toggle is visible in the bottom toolbar; function not yet documented. 3.5 supports both video and image generation; the image-mode picker exposes four Cinematic models (Soul Cinema default, plus Cinematic Characters, Cinematic Locations, and Cinematic Cameras with 2.5 vocabulary) — see higgsfield-cinema for the image-mode surface.
For full Cinema Studio 3.5 documentation → see
higgsfield-cinema
Cinema Studio 4.0 (cinematic_studio_video_4_0) is live as a CLI workflow (not in the MCP
models_explore list, 2026-09-26): the same four modes as Seedance 2.5 (t2v / omni_reference /
video_edit / video_extension), 480p / 720p / 1080p, native audio, plus camera body / lens /
aperture / genre / era / pacing / light / color-palette params. The CLI schema states no duration
bounds — verify before promising a long take. Not yet field-rated; full surface in
higgsfield-cinema § Cinema Studio 4.0.
Related skills
higgsfield-prompt— MCSLA formula, prompt structurehiggsfield-cinema— Cinema Studio model selectionhiggsfield-assist— Credit optimization and plan selectionhiggsfield-audio— Audio-capable model detailshiggsfield-gpt-image-2— GPT Image 2 / 2.5 prompting and when to prefer 2.5templates/— Annotated templates with per-genre model recommendations