Z-Image + Hunyuan Video 1.5 Video Generation Workflow: Image-to-Video Complete Pipeline
From Stills to Motion
Z-Image already produces stunning images, but a static frame always lacks the dimension of time. On November 20, 2025, Tencent's Hunyuan team open-sourced Hunyuan Video 1.5 — a lightweight video generation model with only 8.3B parameters that delivers flagship-quality output on consumer GPUs (24GB VRAM). Based on the official technical report (arXiv 2511.18870), the ComfyUI official blog, and community benchmarks, this article walks through the complete "Z-Image for the frame, Hunyuan Video 1.5 for the motion" image-to-video pipeline.
Why Z-Image + Hunyuan Video 1.5 Is a Golden Combo
Video generation faces two hard problems: first-frame quality and motion coherence.
| Stage | Traditional Approach | This Pipeline |
|---|---|---|
| First frame | T2V model generates it, composition uncontrolled | Z-Image controls composition, lighting, style precisely |
| Motion | Model drifts freely | Hunyuan Video 1.5 anchors to the first frame, preserving character/scene identity |
| Cost | Cloud inference on huge models | Both models run locally; Z-Image Turbo generates in sub-second |
| Control | Text can't specify the exact frame | The image IS the prompt — zero detail loss |
Z-Image Turbo handles "setting the tone" — producing a high-quality first frame in under a second; Hunyuan Video 1.5 handles "making it move" — generating 5–10 second coherent videos from that frame. This is exactly why the two workflows sit side by side in the official ComfyUI template library.
Hunyuan Video 1.5 Core Specs
Model Specifications
| Item | Value |
|---|---|
| Parameters | 8.3B (DiT architecture) |
| Released | 2025-11-20 (weights + code, fully open) |
| Modes | Text-to-video + image-to-video (unified model) |
| Native resolution | 480p / 720p (built-in distilled super-resolution to 1080p) |
| Frame rate | 24 fps |
| Default length | 121 frames (~5 sec); 5/8/10 sec supported |
| Text encoders | Qwen 2.5-VL 7B + byT5 (glyph-aware) |
| Attention | SSTA (Selective and Sliding Tile Attention) |
| VAE | 3D causal VAE |
SSTA is the key innovation: global attention over long video sequences explodes computationally; SSTA applies global attention only to selected tokens and sliding-window local attention to the rest — balancing quality and speed. This is the core reason an 8.3B model runs on consumer hardware.
Capability Highlights
- Strong instruction following: camera movements (pan/tilt/dolly), physics (gravity, collisions), and emotional expressions all respond well to prompts
- In-video text rendering: supports Chinese and English text inside videos — posters, subtitles, and signage can be generated directly
- Diverse styles: realistic, anime, and 3D styles
- Excellent consistency: in I2V mode, preserves the input image's character identity, color palette, and visual style
⚠️ License Note
The Hunyuan Video 1.5 license does not apply in the European Union, United Kingdom, or South Korea (the license agreement is explicitly territory-limited). Check your jurisdiction and distribution scope before commercial use.
ComfyUI Workflow Setup
Model Preparation
Download from HuggingFace and place into the correct directories:
| File | Destination |
|---|---|
| hunyuan_video_1_5_causal_vae.safetensors | models/vae/ |
| hunyuan_video_1_5_llava_llama3_vision.safetensors | models/clip_vision/ |
| hunyuan_video_1_5_qwen2.5_vl_7b_fp8.safetensors | models/text_encoders/ |
| hunyuan_video_1_5_1b_byt5_fp8.safetensors | models/text_encoders/ |
| hunyuan_video_1_5_fp8_e4m3fn.safetensors | models/diffusion_models/ |
If VRAM is tight, select fp8 weight_dtype in the Load Diffusion Model node to significantly reduce memory usage.
Standard I2V Node Graph
[Load Image] → IMAGE (first frame, 1280×720 or 720×1280 recommended)
[Load Hunyuan I2V Model] → MODEL
[CLIP Vision Encode] → image conditioning
[HunyuanVideo15ImageToVideo] → LATENT
[KSampler] (20–30 steps)
[VAE Decode] → [Save Video]
Key point: the first frame strongly drives composition; the prompt drives motion and atmosphere. Use sharp, well-composed inputs whose aspect ratio matches your target output (avoids cropping and letterboxing).
Prompt and Parameter Configuration
Recommended prompt structure (JSON format):
{
"prompt": "A stylish girl walking slowly past a busy restaurant, camera pans left",
"motion": "slow walk, natural arm swing",
"camera": "pan left, slight dolly in",
"aesthetic": "cinematic, shallow depth of field, warm tones"
}
Core parameters:
| Parameter | Recommended | Notes |
|---|---|---|
| Steps | 20 (default) / 30 (high quality) | More steps = better quality, longer render |
| Resolution | 720p (1280×720) | Native output, upscale to 1080p after |
| Frames | 121 (~5 sec) | Extendable to 8/10 sec |
| Image influence | image_token_selection_expr ::4 |
Higher values reduce image influence (::8/::16 weaken it) |
| EasyCache | On (minor quality trade-off) | Caches encodings for repeated runs, big speedup |
Image token selection tip: image_token_selection_expr controls how strongly the image conditions the video. Default ::4 keeps enough constraint; for freer motion, raise it to ::8 or ::16.
Real-World Case and Performance
Case: Cinematic Z-Image First Frame → Animated
- Z-Image Turbo generation: prompt "a rainy neon street at night, a girl in a leather jacket walking" — sub-second, composition and lighting fully controlled
- Hunyuan Video 1.5 animation: load the I2V workflow, upload the first frame, prompt "slow walk, camera pans left, raindrops falling"
- Output: 720p / 5 sec / 24fps video with consistent character identity and color grade throughout
Performance Reference (RTX 4090)
| Stage | Time |
|---|---|
| Generate 5-sec 720p clip (30 steps, Q5_K_M quantized) | ~13 min |
| RIFE 24→60fps interpolation | ~1 min |
| Real-ESRGAN x2 upscale to 1440p | ~4 min |
| Total | ~18 min |
First run downloads model files, adding extra time. TeaCache and similar caching accelerators can push speed further.
Advanced Tips
- Multi-shot storytelling: generate first frames of the same scene from different camera positions (wide/medium/close-up) with Z-Image, animate each, and edit them together for controlled multi-shot narratives
- First/last frame control: some workflows accept both a start and end frame for smoother transitions
- Upscale pipeline: 720p → latent-space upscale → distilled super-resolution → 1080p, a visible quality boost
- Batch production: render multiple candidate first frames with Z-Image Turbo, queue them all through I2V — ideal for short-video asset pipelines
FAQ
Q1: Not enough VRAM?
Load the diffusion model with fp8 weights; drop to 480p; reduce frame count. 24GB runs smoothly; 16GB is worth a try with fp8.
Q2: Character "morphs" or drifts?
Keep the first frame sharp and well-composed; slightly reduce steps to avoid overfitting; keep the prompt concise and non-conflicting with the image.
Q3: Motion too weak?
Raise image_token_selection_expr (e.g. ::8) to loosen image constraint; describe motion explicitly in the prompt.
Q4: Can Hunyuan Video 1.5 be used commercially?
Mind the territory restriction: the license does not apply in the EU, UK, or South Korea. Read the agreement carefully for other regions.
Summary
The Z-Image + Hunyuan Video 1.5 combo seamlessly bridges image generation and video generation: Z-Image Turbo provides a perfectly composed first frame in under a second, and the 8.3B Hunyuan Video 1.5 turns it into a 5–10 second cinematic clip on a 24GB consumer GPU. For short-video creators, ad producers, and indie artists, this is the lowest-cost, most controllable local video production path — from still to motion, just one ComfyUI workflow away.