Skip to content

MODELS.md

Latest commit

 

History

History
123 lines (83 loc) · 10.6 KB

MODELS.md

File metadata and controls

123 lines (83 loc) · 10.6 KB
Raw

Models

WanGP supports video, image, speech, music, and sound-generation models through a common interface. This page is a practical guide to the main model families, not an exhaustive list of every checkpoint, quantization, accelerator, or finetune.

The model selector in WanGP is the source of truth for the models available in your installation. It includes descriptions, model-specific help, and settings suited to each checkpoint. You can also use the toolbar search to switch models quickly.

Clicking a linked model name below loads it directly in WanGP.

Recommended starting points

Goal Suggested model Why pick it
General image generation Krea 2, Z-Image Turbo Krea 2 targets polished, aesthetic images; Z-Image is an efficient 6B model suited to fast iteration.
Image editing and identity preservation Krea 2 Identity Edit, Qwen Image Edit Plus Krea 2 Identity Edit accepts up to two references and includes inpainting/outpainting. Qwen is strong at multi-subject composition and text in images.
Typography and graphic design Ideogram 4 Designed for layout, typography, and high-resolution graphics. Its visual prompt helper can build the model's structured JSON prompt.
Cinematic video with native audio LTX-2.3 Distilled 1.1 Generates video and audio together and supports start/end frames, control video, references, outpainting, injected frames, and sliding windows.
Connected stories with recurring characters JoyAI-Echo Surgical Uses reusable memories across shots to retain characters, voices, objects, and locations.
General text-to-video or image-to-video Wan 2.2 T2V, Wan 2.2 I2V Mature general-purpose video models with broad LoRA and WanGP feature support.
Video editing and outpainting Wan VACE, Bernini-R VACE provides mask/control workflows; Bernini can edit a video or generate from multiple reference images.
Character animation and motion transfer SCAIL-2, Wan 2.2 Animate Animate or replace performers using reference images, pose/control video, and masks.
Talking heads and long dialogue LongCat Avatar 1.5, InfiniteTalk Audio-driven avatar generation with sliding-window support for longer speech.
Voice cloning and dialogue Qwen3 TTS Base, IndexTTS2 Qwen3 is a flexible voice-cloning baseline; IndexTTS2 adds expressive emotion control.
Songs with lyrics ACE-Step 1.5 XL Strong lyric adherence and full-song generation with the XL audio model.
Music and sound effects Stable Audio 3 Generates instrumentals, loops, ambience, and sound effects from descriptive prompts.

Video models

LTX-2 and LTX-2.3

LTX-2 is WanGP's most flexible native audio-video family. Current 22B variants can generate images or videos, synthesize a soundtrack with the visuals, and work with:

  • text, start/end frames, injected frames, control video, and control audio;
  • video continuation and sliding windows;
  • pose, depth, edge, HDR, inpainting, outpainting, and other IC-LoRA processes;
  • talking-head generation and voice conditioning;
  • multi-phase or direct high-resolution generation, depending on the selected model and process.

Use LTX-2.3 Distilled 1.1 as the fast general-purpose starting point. Dev variants trade speed for more control. LTX-2.3 MSR V2 accepts two to five subject/object references, while JoyAI-Echo Surgical is intended for connected multi-shot stories.

Wan 2.1 and Wan 2.2

The Wan family includes general generation, editing, control, motion-transfer, identity, and talking-head architectures:

  • Wan 2.2 T2V and Wan 2.2 I2V for general generation;
  • Wan 2.2 TI2V 5B for a smaller unified text/image-to-video path;
  • VACE for inpainting, outpainting, object replacement, motion/depth/pose control, and other video-editing workflows;
  • Bernini-R for video-to-video generation and multi-reference conditioning;
  • SCAIL-2 and Wan 2.2 Animate for character animation or performer replacement;
  • Lynx and VACE Lynx for identity-preserving face replacement;
  • MultiTalk and InfiniteTalk for audio-driven dialogue;
  • Vista4D for reshooting a dynamic scene from a new camera trajectory.

WanGP also ships accelerated defaults and specialist finetunes. Treat names such as FusioniX, FastWan, Lightning, Self-Forcing, Cocktail, Sparse, and NVFP4 as variants of a base workflow rather than separate model families.

Other video families

  • Kandinsky 5 Pro provides 19B text-to-video and image-to-video models with controllable camera motion. Lite and distilled variants reduce generation cost.
  • HunyuanVideo 1.5 provides 8.3B text-to-video and image-to-video models, plus distilled and upsampler variants.
  • LongCat Video is a general video model; LongCat Avatar 1.5 is its distilled audio-driven avatar model.
  • Ovi generates video with a synchronized soundtrack and is especially suited to speaking characters.
  • Magi Human generates audio-driven talking-head video and offers staged high-resolution variants.
  • Legacy families such as HunyuanVideo, LTX-Video 13B, Phantom, Recam, FantasySpeaking, SkyReels Diffusion Forcing, and Wan-Fun remain available where useful.

Image models

Krea 2

Krea 2 RAW is the undistilled CFG-guided checkpoint; Krea 2 Turbo is the faster distilled choice. Identity Edit variants add instruction-based editing with up to two reference images. WanGP also exposes inpainting through LanPaint and negative prompting on distilled Krea 2 through NAG.

Qwen Image

Qwen Image includes text-to-image, image editing, multi-reference editing, and layered-image variants. Qwen Image Edit Plus is a strong default for combining subjects and objects, preserving a scene while editing it, and rendering longer text. Quantized Nunchaku variants are available for supported hardware.

Ideogram 4

Ideogram 4 focuses on typography, layout, graphic design, and structured composition. Plain prompts work, but the model-specific Magic Prompt and visual helper make it easier to author its JSON format. Turbo Time and NF4 variants provide alternative speed or memory trade-offs.

Flux, Z-Image, and HiDream

  • Flux includes Schnell, Dev, Kontext, Krea, Chroma, SRPO, Flux 2, and Klein architectures for generation, editing, and reference-guided workflows.
  • Z-Image Turbo is an efficient distilled 6B generator. Base, Control, Control 2.x, TwinFlow, and Nunchaku variants extend it to editing and lower-memory execution.
  • HiDream O1 supports text-to-image and image-reference generation, with control-image and latent-preview integration in WanGP.

Speech, music, and sound

WanGP groups audio-only models in the same selector:

  • Qwen3 TTS Base supports reference-audio voice cloning and two-speaker dialogue. Custom Voice and Voice Design are separate variants.
  • IndexTTS2 supports zero-shot voice cloning, long dialogue, and text- or audio-guided emotion.
  • OmniVoice supports multilingual speech, voice design, cloning, and dialogue.
  • Chatterbox and KugelAudio provide alternative speech and cloned-dialogue workflows.
  • MiniMax Music 3 generates complete lyrics- and description-conditioned songs up to five minutes with native 44.1 kHz stereo output. Its flow transformer and 8B autoregressive Qwen text encoder use persistent ConvRot INT8 checkpoints, while its RVQ decoder has a dedicated accelerated checkpoint and all placement remains MMGP-managed. The shared lm_decoder_engine setting exposes CUDA-graph (cg) decoding and a vllm path that is selected only when both FlashAttention2 and Triton are available.
  • ACE-Step 1.5 XL and HeartMuLa generate songs, including lyrics-driven tracks.
  • Stable Audio 3 generates music, loops, ambience, and sound effects.
  • Scenema Audio and DramaBox use LTX-2's audio knowledge for expressive scene-aware speech and dialogue.

Audio postprocessing is model-independent. For example, SeedVC can replace one or two voices in an existing audio or video result, while MMAudio and PrismAudio can add sound to silent video.

Choosing variants

Model size alone does not determine whether a checkpoint will run. Resolution, frame count, attention backend, quantization, memory profile, tiling, sliding-window size, and enabled controls all affect RAM and VRAM use. The old fixed “6 GB / 12 GB / 20 GB” model tiers are therefore not reliable.

For a new workflow:

  1. Start with the recommended or distilled variant.
  2. Select a memory profile appropriate for your GPU.
  3. Test a short clip or smaller image before increasing resolution or duration.
  4. Use sliding windows for supported long-video workflows.
  5. Add LoRAs, controls, references, and upsampling one feature at a time.

Quantized defaults such as quanto int8, FP8, GGUF, NVFP4, NF4, and Nunchaku reduce memory or target a specific backend. They should preserve the model's workflow, although speed, compatibility, and quality can vary with hardware and kernels.

Model switching and custom models

WanGP loads models on demand and can switch without restarting. Use the model selector or toolbar search; the previous model is unloaded as needed to recover memory. Saved settings can retain model inputs and generation options.

The model list can be refreshed after adding or editing a finetune. User-provided checkpoints belong in finetunes/; model plugins can contribute additional families and defaults. See Finetunes, LoRAs, and Plugins for details.