Imported from osidemedia/higgsfield-ai-prompt-skill (skills/higgsfield-models/SKILL.md). Install upstream with npx skills add osidemedia/higgsfield-ai-prompt-skill --skill higgsfield-models. Copyright stays with the author.

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.

Quick Decision Flowchart

Fast lookup — for detailed comparisons see the full tables below.

Pricing tiers: Free (Soul 2.0, DoP Lite) · 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

Decision Flowchart

Is this image or video?
├── IMAGE
│   ├── Person / portrait? → Soul 2.0
│   ├── Cinematic keyframe for I2V pipeline? → Soul Cinema Preview
│   ├── Native 4K / image series / storyboarding? → Kling Image 3.0
│   ├── 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 1.5
│   └── Edit an existing image? → Flux Kontext
│
└── VIDEO
    ├── EDIT existing footage?
    │   ├── Relight, restyle, swap, remove → Kling O1 Video Edit
    │   └── Higher quality 3.0 edit → Kling 3.0 Omni Edit
    │
    ├── Is a human character the focus?
    │   ├── Need audio, long clip (15s), multi-shot → Kling 3.0
    │   ├── Need to clone from reference video → Kling 3.0 Omni
    │   ├── Best lip-sync + multilingual → Seedance 1.5 Pro
    │   ├── Legacy-tier great character (audio togglable via `sound`) → Kling 2.6
    │   └── Fast iteration → Kling 2.5 Turbo
    │
    ├── Need motion transfer from reference video?
    │   └── → Kling 3.0 Motion Control
    │
    ├── Animate a still image with cinematic camera?
    │   └── → Higgsfield DoP (Lite/Standard/Turbo)
    │
    ├── Is the environment/phenomenon the hero?
    │   ├── Nature, documentary, stable → Veo 3
    │   ├── Need ref image consistency → Veo 3.1
    │   ├── Budget Veo 3.1 quality / volume → Veo 3.1 Lite
    │   ├── 60fps, first+last frame, ref images → Wan 2.7
    │   └── Artistic, painterly, fantasy → Wan 2.5/2.6
    │
    ├── Is it action/spectacle?
    │   ├── Epic scale, crowds, physics → Sora 2
    │   ├── VFX, anime, fluid motion → Minimax Hailuo 2.3
    │   └── Dance, sports, budget motion → Minimax Hailuo 02
    │
    ├── Need maximum reference control?
    │   ├── Up to 12 assets (images+video+audio) → Seedance 2.0
    │   ├── Up to 7 image refs → Kling O1 Video
    │   ├── Image + video refs with native audio (4–10s, 720p) → Gemini Omni Flash
    │   └── Up to 3 asset refs → Veo 3.1
    │
    └── Speed/cost priority?
        ├── Fastest Kling → Kling 2.5 Turbo
        ├── Fastest Seedance → Seedance Pro
        └── Fastest Veo → Veo 3.1 Fast or Veo 3 Fast

Image Models — Quick Selection

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): Higgsfield Soul · Kling O1 · Seedream family · Nano Banana · Wan 2.2 · Reve
• Mid (1.5–2): Nano Banana 2 · Multi Reference · FLUX.2 Pro · Flux Kontext Max · GPT Image · NB Pro · Character Swap
• Premium (5–6): FLUX.2 Flex · FLUX.2 Max

General pricing tiers (video + image, approximate):

• Free: Soul 2.0 · DoP Lite (limited)
• 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

Key Model Notes

Kling 3.0 vs 2.6: 3.0 is the current top 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.

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).

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).

Veo 3.1 vs 3.1 Lite vs 3: 3.1 adds reference images (up to 3), first/last frame, video extension, 4K. 3.1 Lite is budget-priced 3.1 quality at 1080p — supports T2V and I2V, 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.

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 remains available for budget motion work.

Grok Imagine: Aurora architecture (autoregressive, not diffusion) — excels at text/logo rendering and multi-image compositing. Image editing supports multi-turn iterative chains.

Gemini Omni Flash: Google's reference-driven video model — accepts both image and video references, generates native audio, 4–10s at 720p (16:9 or 9:16). Use it when you want reference-anchored clips with sound at a fast/light tier rather than flagship fidelity.

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.

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

Related skills

• higgsfield-prompt — MCSLA formula, prompt structure
• higgsfield-cinema — Cinema Studio model selection
• higgsfield-assist — Credit optimization and plan selection
• higgsfield-audio — Audio-capable model details
• templates/ — Annotated templates with per-genre model recommendations