logoChat01.ai

https://github.com/Tencent-Hunyuan/HunyuanVideo-Av...

Created on: June 11, 2025

Answered using O3 by Chat01

Question

https://github.com/Tencent-Hunyuan/HunyuanVideo-Avatar

Is there any API for this model in repicate/fal.ai or alike platforms, if not, is there any other model API to make one Avator image to video?

Thought

Clarifying Hunyuan Avatar Availability

I couldn't find an official "hunyuanAvatar" on Replicate. The search led to custom fine-tunes, like lucataco/hunyuan-heygen-woman, rather than the video avatar. So, Replicate doesn't host the official version. I gathered more details from other models like "talking-head-1" on Replicate. To clarify: there is no official HunyuanVideo-Avatar API on Replicate, but Fal.ai does host it as "hunyuan-avatar" with an accessible API. I'll note that as my final answer.

Clarifying Hunyuan Avatar Availability

Short answer: yes—Fal.ai already offers an official HunyuanAvatar API. Replicate hosts the text-to-video HunyuanVideo but not the Avatar variant at the moment.

Alternatively, other services offer image-to-video APIs such as SadTalker, VideoReTalking, AnimateDiff, and HeyGen, which can generate avatar videos. There are also some additional models available on Fal.ai like PixVerse and Kling.

As of June 11, 2025, we can rely on various sources, including official docs, to gather detailed information.

Thought for 1m 50s

Answer

Short answer: Fal.ai already exposes an official fal-ai/hunyuan-avatar endpoint that turns a single avatar image + audio into a talking-head video. Replicate, by contrast, only hosts the text-to-video model tencent/hunyuan-video and a few unofficial fine-tunes, so there is no first-party HunyuanVideo-Avatar API there. If you just need an “image-to-video avatar” pipeline, several other public APIs—SadTalker, VideoReTalking, AnimateDiff, HeyGen, A2E, and Fal.ai’s Kling or PixVerse—can fill the gap.

1 ▪ Status of HunyuanVideo-Avatar on popular inference hubs

PlatformOfficial Avatar model?Notes
Fal.aiYes – fal-ai/hunyuan-avatarAccepts one image & audio; playground + REST/SDK; ~8 min/5 s video, about $1.40 per run fal.ai
ReplicateNo. Only tencent/hunyuan-video (text→video) and community fine-tunestencent/hunyuan-video ≈104 k runs, full API; but it requires a text prompt, not an avatar image replicate.com. A proof-of-concept fine-tune lucataco/hunyuan-heygen-woman exists, but it is not the official Avatar model replicate.com

2 ▪ Grab-and-go APIs that animate a single image

Model / ProviderWhat it doesTypical inputsCost / speed (public default)
SadTalker – ReplicateAudio-driven talking-head from one photoimage + audio≈$0.19 / ~2 min for 1 clip replicate.com
VideoReTalking – ReplicateLip-sync or re-voice an existing face video; also works from one stillimage or video + audio≈0.26percall,3runs/0.26 per call, 3 runs / 1 replicate.com
AnimateDiff – ReplicateAdds short looped motion to a still image (no audio)image (+prompt)≈$0.013, ~10 s per 16-frame clip replicate.com
Kling 2.1 / PixVerse / Veo 3 – Fal.aiHigh-quality image-to-video (no audio) with realistic motionimage (+prompt)Tiered ($) endpoints shown in Fal gallery fal.ai
HeyGen APICommercial avatar generator; create, script, translate & lip-sync videosimage/photo avatar + text or audioUsage-based pricing; full REST API (/avatar-video) docs.heygen.com docs.heygen.com docs.heygen.com
A2E Image-to-Video APIEnd-to-end avatar, lip-sync, voice-cloneimage + audio or textFree tier + paid plans; REST endpoints a2e.ai

3 ▪ Choosing the right option

  1. Exact model parity: If you specifically need HunyuanVideo-Avatar, use Fal.ai today—deployment is turnkey and follows the repo’s audio-driven pipeline.
  2. Replicate-only environment: Combine SadTalker for lip-synced faces or VideoReTalking for re-voicing; for silent motion, AnimateDiff works well.
  3. Higher-fidelity cinematic motion: Fal.ai’s Kling 2.1/PixVerse or Veo 3 endpoints deliver 1080 p+ footage, though you must overlay audio separately.
  4. All-in-one SaaS: Platforms like HeyGen or A2E bundle voice cloning, subtitle translation and template scenes—ideal when you want production-ready videos without managing ML infra.
  5. Self-hosting: You can still clone the GitHub repo and run locally, but you’ll need multi-GPU compute and to expose your own inference endpoint—Fal.ai’s hosted API is usually simpler unless you require full control.

Quick start with Fal.ai’s endpoint

bash
curl -X POST https://fal.run/fal-ai/hunyuan-avatar \
  -H "Authorization: Key $FAL_API_KEY" \
  -F "image_url=https://example.com/avatar.jpg" \
  -F "audio_url=https://example.com/voice.mp3" \
  -F "emotion=happy" \
  -F "fps=25"

The response JSON contains an output_with_audio.mp4 URL—download or stream in your app fal.ai.

4 ▪ Future watch-list

  • Tencent’s team released the avatar paper on May 26 2025; public model ports evolve rapidly, so an official Replicate wrapper may appear soon arxiv.org.
  • Replicate announced a video playground that simplifies chaining image→video models and could host more avatar checkpoints in coming months replicate.com.

TL;DR

Use Fal.ai’s hunyuan-avatar if you want the exact model right now. Otherwise, pick from Replicate (SadTalker, VideoReTalking, AnimateDiff) or commercial APIs like HeyGen and A2E depending on whether you need lip-sync, emotion control, or large-scale production. All provide REST endpoints you can integrate in minutes.

Question

It has been some time, please try to search again, and this time, add https://github.com/deepbeepmeep/Wan2GP too

Thought

Searching for Wan2GP updates

I’ve found quite a few results while searching for Wan2GP model updates, including the GitHub repository. The model has endpoints available on Fal (wan-pro, image-to-video) and Replicate (wan-2.1 models). Some models might have direct connections, but it seems there aren’t many clear vantage points available for direct access. I’ll summarize these updates to ensure I can provide a concise answer, focusing on the potential applications and key platforms.

Summarizing findings and candidate solutions

As of June 14, 2025:

  • Fal.ai still has the Hunyuan-avatar endpoint unchanged, while Replicate lacks an official Hunyuan avatar but hosts fine-tuned versions. No new official updates for Replicate's Hunyuan avatars.
  • Wan2GP started as a GitHub project to run Wan 2.1 on consumer GPUs. Endpoints for Wan 2.1 variants now exist on Replicate and Fal.ai.
  • Other solutions like SadTalker, VideoReTalking, Luma Dream Machine Video, and MiniMax are also available.
  • Example code for Replicate or Fal endpoints will be provided, along with multiple citations for each point.

With that, I'll make sure to structure the response as required with citations after each section.

Thought for 27 seconds

Answer

TL;DR (mid-June 2025):
Fal.ai is still the only public hub that exposes an official API for HunyuanVideo-Avatar (fal-ai/hunyuan-avatar). fal.ai Replicate has not added a first-party Avatar endpoint; it only offers the core text-to-video model (tencent/hunyuan-video) and a few hobby finetunes. replicate.com replicate.com
By contrast, the Wan-2.1 / Wan2GP ecosystem now has multiple hosted options: Fal.ai runs several image-to-video endpoints (wan-i2v, wan-pro/image-to-video, etc.) fal.ai fal.ai, and Replicate lists high-resolution forks such as wavespeedai/wan-2.1-i2v-720p and LoRA-enabled variants. replicate.com replicate.com The GitHub repo Wan2GP itself remains a self-host kit (10 GB VRAM) rather than a web API, but its checkpoints are binary-compatible with these services. github.com github.com

1 HunyuanVideo-Avatar: where you can call it today

PlatformEndpoint IDStatusInputsTypical cost
Fal.aifal-ai/hunyuan-avatarOfficialimage_url, audio_url, optional emotion, fps~$1.40 / 5 s clip fal.ai
Replicatenone (core text-to-video only)––– replicate.com
Replicate (community)lucataco/hunyuan-heygen-* fine-tunesunofficialstill + prompthobby forks replicate.com
Hugging Facetencent/HunyuanVideo-Avatarweights onlylocal inferencefree download huggingface.co

Quick Fal.ai call

bash
curl -X POST https://fal.run/fal-ai/hunyuan-avatar \
  -H "Authorization: Key $FAL_API_KEY" \
  -F image_url=https://…/face.jpg \
  -F audio_url=https://…/voice.mp3 \
  -F emotion=happy -F fps=25

The JSON response includes an output_with_audio.mp4 URL. fal.ai

2 Wan-family models (Wan-2.1 & Wan2GP)

2.1 Fal.ai options

  • fal-ai/wan-i2v – 480 p/720 p image-to-video, $0.20–$0.40 per clip fal.ai
  • fal-ai/wan-pro/image-to-video – premium 1080 p, 6 s, 30 fps endpoint fal.ai
  • Training: fal-ai/wan-trainer lets you fine-tune LoRAs on your own clips. fal.ai

2.2 Replicate options

  • wavespeedai/wan-2.1-i2v-720p – optimized 720 p generation replicate.com
  • fofr/wan2.1-with-lora – LoRA-ready build, H100 backend replicate.com
  • Replicate’s blog tutorial shows how to call any Wan-2.1 variant with a single line of code. replicate.com

2.3 Where Wan2GP fits

The GitHub repo packages Wan-2.1 weights plus a TeaCache patch so you can run 720 p clips on a single 10 GB consumer GPU. github.com It also documents how to export Replicate-style LoRAs for cloud inference. github.com If you prefer an HTTP API, point your client to the Fal.ai or Replicate endpoints above—both accept the same latent formats.

3 Still-image-to-video avatar alternatives

Model/APIHubStrengthCitation
SadTalkerReplicateAudio-driven talking head from one photoreplicate.com
VideoReTalkingReplicateLip-syncs or re-voices still or videoreplicate.com
HeyGen /avatar-videoSaaSCommercial studio-grade avatar videosdocs.heygen.com
PixVerse / Kling 2.xFal.aiHigh-motion silent clips (image-to-video)fal.ai

These remain the main drop-in options if the Hunyuan or Wan stacks don’t meet your quality, licensing, or cost targets.

4 Choosing between them

  1. Exact HunyuanAvatar pipeline → use Fal.ai’s managed endpoint.
  2. Higher-resolution cinematic motion without audio → Wan-Pro, PixVerse, or Kling on Fal.ai.
  3. Replicate-only environments → Wan-2.1 forks for motion, plus SadTalker/VideoReTalking for lip-sync.
  4. Self-hosting → Clone Wan2GP or HunyuanVideo-Avatar from GitHub and run locally; you can later containerise with Cog to deploy on Replicate. github.com github.com
  5. Turn-key SaaS → HeyGen if you need an all-in-one script-to-video pipeline.

Need code examples, pricing calculators, or a deeper comparison?

Let me know and I can produce ready-to-run snippets or benchmark tables.

Share this Q&A