HunyuanVideo-I2V — Public project availability¶

Development line: project:hunyuanvideo-i2v · thread public-project-availability
Last event: 2025-03-06 · 1 dated since 2025-03-06 · Researched: 2026-09-04 · confidence: medium

What it is¶

HunyuanVideo-I2V is Tencent’s 2025 local image-to-video stack for teams that need first-frame-guided generation rather than a hosted API.

• Inputs: reference image and text prompt conditioning.
• Components: official PyTorch code, pretrained weights, sampling, LoRA training, and xDiT multi-GPU inference.
• Output: video generation up to 720p and 129 frames (about five seconds).

Official single-GPU guidance specifies 60 GB peak VRAM and tests on 80 GB.

The original stack remains runnable, but new deployments should compare the later HunyuanVideo-1.5 I2V line before committing hardware and integration work.

Development line¶

• 2025-03-06 — HunyuanVideo-I2V public project release. Tencent released the inference code and model weights.

What changed¶

HunyuanVideo-I2V starts with a local 720p release, corrects its initial weights, and adds multi-GPU inference.

• 2025-03-06: Tencent released the inference code and model weights.
• 2025-03-07: Tencent replaced the release state after fixing an ID-change bug; the corrected weights are the source of truth for first-frame consistency.
• 2025-03-09: Tencent published no separate official code, weights, or version change dated this day; no implementation delta can be verified.
• 2025-03-11: Tencent updated LoRA training and inference code after the bug fix.
• 2025-03-13: Tencent released xDiT-powered parallel inference.
• 2025-11-21: Tencent released the separate HunyuanVideo-1.5 T2V/I2V line, with 480p and 720p I2V checkpoints; it is not documented as a drop-in continuation of this repository.

How to use this¶

From 2025-03-06, practitioners should use the linked official Hugging Face page and Tencent GitHub repository as the canonical starting points for HunyuanVideo-I2V, and wait for archived update contents before acting on later announcements.

1. Confirm the original stack fits the machine: Linux and an NVIDIA CUDA GPU are required; plan for 60 GB peak VRAM at 720p, with 80 GB the tested configuration. — https://github.com/Tencent/HunyuanVideo-I2V?tab=readme-ov-file#-news
2. Clone the official repository, create its Python 3.11.9 environment, and install the documented PyTorch/CUDA dependencies and requirements; use the published CUDA 12 container if that matches the deployment. — https://github.com/Tencent/HunyuanVideo-I2V?tab=readme-ov-file#-news
3. Download the current official HunyuanVideo-I2V weights from the Tencent model card and review the Tencent Hunyuan Community License before use. — https://huggingface.co/tencent/HunyuanVideo-I2V
4. Run sample_image2video.py with HYVideo-T/2, --i2v-mode, a reference-image path, a concise prompt, 720p resolution, sampling steps, video length, and an output path; enable CPU offload when needed for high-resolution generation. — https://github.com/Tencent/HunyuanVideo-I2V?tab=readme-ov-file#-news
5. Choose stable or dynamic motion deliberately: stable mode uses --i2v-stability with --flow-shift 7.0; dynamic output uses stability off with --flow-shift 17.0. Set a seed when the result must be reproducible. — https://github.com/Tencent/HunyuanVideo-I2V?tab=readme-ov-file#-news
6. For a custom effect, prepare video-caption training data, run the supplied LoRA training script, then load the produced safetensors file through --use-lora and --lora-path during inference. — https://github.com/Tencent/HunyuanVideo-I2V?tab=readme-ov-file#-news

Best practices¶

• Do not retain the March 6 launch weights: use the corrected weights announced after the identity-change fix. — https://github.com/Tencent/HunyuanVideo-I2V?tab=readme-ov-file#-news
• Keep the prompt short and concrete: specify subject and action, then add background or camera only when needed; overly detailed prompts can introduce unwanted transitions. — https://github.com/Tencent/HunyuanVideo-I2V?tab=readme-ov-file#-news
• Use flow shift 7 with stable mode for steadier motion, and flow shift 17 with stable mode off when motion matters more than stability. — https://github.com/Tencent/HunyuanVideo-I2V?tab=readme-ov-file#-news
• Budget hardware from the documented 60 GB minimum rather than assuming CPU offload makes the 720p workflow low-VRAM; the official tested setup is 80 GB. — https://github.com/Tencent/HunyuanVideo-I2V?tab=readme-ov-file#-news
• For LoRA effects, put a phrase or short-sentence trigger directly in each video caption; train at 360p, then use the resulting LoRA for 720p inference. — https://github.com/Tencent/HunyuanVideo-I2V?tab=readme-ov-file#-news

Superseded by this¶

• 2025-03-06 initial HunyuanVideo-I2V weights are superseded by the 2025-03-07 corrected weights because the initial open-source release could change identity.
• The 2025-03-06 deployment choice is dated, not formally deprecated: HunyuanVideo-1.5 released on 2025-11-21 as a separate later Tencent T2V/I2V line with 480p and 720p I2V checkpoints. Evaluate it for new deployments, but do not assume checkpoint or workflow compatibility.

Still unknown¶

• The March 9 event cannot be tied to a distinct upstream release from first-party records; the linked source page was not retrievable.
• Tencent does not document HunyuanVideo-1.5 as a drop-in upgrade or formal replacement for HunyuanVideo-I2V.
• The sources state no current maintenance commitment, support window, or hosted API availability for the original HunyuanVideo-I2V stack.

Sources¶

Agent brief¶

• Subject: project:hunyuanvideo-i2v, thread public-project-availability, 1 dated events 2025-03-06 → 2025-03-06.
• Practical note: From 2025-03-06, practitioners should use the linked official Hugging Face page and Tencent GitHub repository as the canonical starting points for HunyuanVideo-I2V, and wait for archived update contents before acting on later announcements.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.