AccVideo vs Hunyuan: Faster AI Video in ComfyUI

I’ve been testing a new approach for speeding up Hunyuan Video, and the results are very promising. This method doesn’t rely on ultra-low sampling steps like LCM tricks seen elsewhere. In my runs, quality stayed at the same level as standard Hunyuan Video, and in some cases improved.

The key is a model called AccVideo. It’s presented as a video fusion approach trained with synthetic datasets, designed to generate Hunyuan-style clips much faster. The claim is 8.5x speed-up, and based on my tests, that figure holds up at 720p.

15 steps: 3-second clip in about 3 minutes
15 steps: 5-second clip in 2 minutes 46 seconds
5 steps: 5-second clip in 1 minute 55 seconds

I also created a mix of cinematic and fantasy-style clips. Motion and character stability were strong, with results in line with Hunyuan Video’s typical look and feel, but with far shorter wait times.

What AccVideo Is?

AccVideo is released as a companion to Hunyuan Video and runs with the same text encoder and VAE. It builds on Hunyuan Video’s text-to-video weights and targets 720p generation as a baseline. You can run it directly in Python, and it’s already easy to integrate into ComfyUI.

The model is available in multiple formats:

As a LoRA for the Hunyuan base
As a quantized GGUF model
As FP8 safetensors for direct loading

This flexibility makes it simple to experiment and compare speed and quality trade-offs across setups.

Why It’s Faster?

AccVideo relies on synthetic training data and model distillation. Distillation is a teacher–student approach that removes extra computation while preserving the core behavior of the original model. In practice, it reduces the sampling burden and shortens inference time without degrading output.

On 720p text-to-video, AccVideo consistently finishes generations many times faster than standard Hunyuan Video under similar conditions. The speed advantage becomes clear as you raise frame count and resolution.

Files and Model Formats

AccVideo is included in the provider’s Hugging Face repository with several download options. They all work with the same Hunyuan text encoder and VAE, so setup is familiar if you’ve already run Hunyuan Video.

Here’s what matters most:

AccVideo T2V FStep (FP8 safetensors)
- Size: about 13 GB
- Purpose: Direct loading as a full model
- Quality: Comparable to quantized Q8 GGUF in most tests
AccVideo GGUF quantizations
- Recommended: Q8 for best balance of size and fidelity
- Advantage: Lower VRAM usage; fast inference
AccVideo LoRA for Hunyuan Video (5-step, rank 16, FP8)
- Size: about 174 MB
- Use case: Apply on top of the Hunyuan Video 720p base model
- Advantage: Minimal storage impact; easy to slot into existing Hunyuan workflows
Optional: Hunyuan Video Reward MPS LoRA
- Purpose: Improves overall quality and stability
- Benefit: Better character consistency across frames

Quick Comparison

Format	Typical Size	Where It Loads	Speed	Quality	Notes
AccVideo LoRA (5-step, FP8)	~174 MB	Hunyuan base via LoRA Loader	Fast	Strong	Easiest drop-in for existing Hunyuan setups
AccVideo GGUF (Q8)	Varies	Direct GGUF Model Loader	Fastest	Strong	Light on VRAM, direct model swap
AccVideo FP8 (safetensors)	~13 GB	Model Loader	Fast	Strong	Closest to full-precision behavior

How AccVideo Speeds Up Hunyuan Video?

In Hunyuan Video, sampling steps drive most of the runtime cost. AccVideo’s training process compresses the behavior of a larger model into a faster student that needs fewer steps to converge.

The result: fewer steps, shorter runs
At 720p, speed-ups of around 8x are common
Quality remains on par with the base model’s look and motion

In practice, you can work with 5–15 steps instead of higher counts, especially once you’ve dialed in your prompt, frames, and sampler settings.

Running AccVideo in ComfyUI

You can run AccVideo in ComfyUI in two main ways:

Apply the AccVideo LoRA on top of Hunyuan Video 720p
Load the AccVideo model directly via GGUF or FP8 files

Both paths are straightforward. The LoRA route is best if you already have Hunyuan Video set up. The GGUF path is the quickest for maximum speed.

Method 1: Hunyuan Video + AccVideo LoRA

This method keeps your standard Hunyuan Video workflow and adds AccVideo as a LoRA. It’s a native setup with good stability.

Steps:

Load Hunyuan Video T2V 720p as your base model.
Add the Hunyuan Video LoRA Loader (double blocks).
Connect the AccVideo 5-step LoRA (FP8) in the first block.
Optionally add the Hunyuan Reward MPS LoRA in the second block for improved quality and consistency.
Wire the LoRA output to your sampler as usual.
Use the same Hunyuan text encoder and VAE.
Set your sampling steps (start with 10), frames (e.g., ~97), and prompt.
Generate.

Notes:

“Double blocks” lets you stack multiple LoRAs cleanly.
This route preserves Hunyuan’s native behavior while accelerating generation.
In my test with 10 steps and 97 frames, inference time was about 1 minute 56 seconds.

Method 2: Direct AccVideo Model (GGUF or FP8)

This method loads AccVideo itself and skips the base Hunyuan model. GGUF is the fastest option and lighter on VRAM.

Steps:

Add the GGUF Model Loader (or FP8 Model Loader).
Load the AccVideo GGUF Q8 (or AccVideo FP8) model.
Bypass the Hunyuan diffusion model loader and the LoRA loader.
Connect the model loader directly to your sampler.
Keep the same text encoder and VAE used by Hunyuan.
Choose your sampling steps (5–15), frames, and prompt.
Generate.

Notes:

Expect very fast runs with GGUF Q8.
You can switch between GGUF and FP8 to compare memory use and quality.

Settings That Worked Well

AccVideo responds well to low and moderate sampling steps. Here’s what I saw across multiple runs.

Steps
- 5 steps: Lowest generation time; still coherent motion and clean frames in many prompts
- 10 steps: My preferred balance for quality and speed
- 15 steps: Slightly smoother motion; longer time but still fast
Frame count
- 90–120 frames is a practical daily range at 720p
- Larger counts scale predictably; AccVideo’s speed helps keep runs under a few minutes
Resolution
- 720p is the primary target for these models
- Higher resolutions add time; start at 720p for tuning
Sampler options
- Standard samplers used in Hunyuan workflows work well
- Keep CFG scale conservative to avoid jitter and drift
Seeds
- Random seeds can introduce motion variance
- If character stability matters, use a fixed seed and the Reward MPS LoRA

Test Timings and Observations

I recorded a consistent reduction in generation time for typical short clips at 720p.

3-second clip at 15 steps: ~3 minutes
5-second clip at 15 steps: ~2 minutes 46 seconds
5-second clip at 5 steps: ~1 minute 55 seconds
97 frames at 10 steps (LoRA route): ~1 minute 56 seconds

Across tests, motion quality matched what I normally expect from Hunyuan Video. Character consistency was improved by stacking the Reward MPS LoRA with the AccVideo LoRA. The GGUF Q8 route was the quickest overall and a strong option when you’re aiming for short iteration cycles.

Generation time varies across prompts. This is normal for a distilled model that is predicting scene dynamics from text. Once you settle on your prompt and frames, timings stabilize.

Why Use the LoRA Route vs. Direct AccVideo

Both methods work, but they serve different needs.

LoRA on Hunyuan base
- Best for those who already have Hunyuan Video running
- Simplifies swapping in other Hunyuan LoRAs (e.g., character LoRAs)
- Keeps the native Hunyuan node structure in ComfyUI
- Easy to stack Reward MPS and AccVideo together
Direct AccVideo (GGUF or FP8)
- Fastest end-to-end, especially with GGUF Q8
- Lower VRAM compared to full-precision models
- Fewer moving parts in the graph

If you’re starting fresh and want speed right away, go straight to the GGUF build. If you’re invested in Hunyuan’s LoRA ecosystem, the LoRA route gives you more flexibility.

Optional Optimizations That Also Work

I tested several optimizations commonly used with Hunyuan Video. They worked with the AccVideo GGUF models and helped shave off extra time while slightly improving quality.

Stage Attention
- Helps with consistent attention across frames
- Can reduce flicker in some prompts
torch.compile control models
- Speeds up parts of the pipeline
- Works with the GGUF route in my tests
Model patcher order adjustments
- Small quality gains
- Useful when stacking multiple tweaks

These are incremental improvements. AccVideo provides the major speed win; these extras polish the output and reduce small inefficiencies.

Step-by-Step: Full ComfyUI Workflow (LoRA Route)

Use this when you want to keep the Hunyuan base and add AccVideo acceleration.

Base Model
- Load Hunyuan Video T2V 720p in the model loader.
Text Encoder and VAE
- Use the standard Hunyuan text encoder and VAE nodes.
LoRA Loader (Double Blocks)
- Insert the Hunyuan Video LoRA Loader with double blocks.
Attach LoRAs
- Block A: AccVideo 5-step LoRA (FP8)
- Block B (optional): Hunyuan Reward MPS LoRA
Connect to Sampler
- Wire the LoRA output to your sampler node.
Set Parameters
- Steps: Start with 10
- Frames: Try around 90–120
- CFG scale: Keep moderate to avoid overshooting
Output
- Decode with the VAE and write frames to video.
Generate
- Monitor timing; you should see a clear speed-up compared to standard Hunyuan.

Step-by-Step: Full ComfyUI Workflow (Direct AccVideo, GGUF)

Use this when speed and simplicity are top priorities.

Model Loader
- Add a GGUF Model Loader node.
Load AccVideo
- Select the AccVideo Q8 GGUF model.
Bypass Diffusion and LoRA Loaders
- Connect the GGUF model directly to the sampler.
Text Encoder and VAE
- Keep the Hunyuan encoder and VAE nodes as usual.
Parameters
- Steps: 5–15 (try 10)
- Frames: 90–120 to start
Output
- Decode and save to video.
Generate
- Expect fast turnaround, typically 1–2 minutes for short clips.

Practical Tips for Stable Results

Start at 720p
- It’s the primary target for these models and the best way to tune prompts and steps.
Use fixed seeds for consistency
- This helps evaluate changes to steps and LoRAs.
Stack LoRAs thoughtfully
- The Reward MPS LoRA pairs well with the AccVideo LoRA to stabilize motion and characters.
Keep CFG modest
- High CFG can produce jitter and oversaturation in motion.
Use quantization wisely
- Q8 GGUF preserved quality in my tests while keeping memory low.
Batch your prompts
- Run a few variations back-to-back to see how timings and motion patterns change. AccVideo’s speed makes iteration practical.

Quality and Stability

In side-by-side runs, AccVideo kept the visual style I associate with Hunyuan Video: clean shading, solid motion, and coherent sequences. Even at low step counts, frames lined up well, and transitions held together through typical camera or subject motion.

Adding the Reward MPS LoRA improved shot-to-shot stability and reduced character drift, especially on longer frame counts. It’s a useful pair with AccVideo when consistency matters.

Storage and VRAM Notes

LoRA route
- Minimal storage footprint (the AccVideo LoRA is ~174 MB)
- Uses the Hunyuan base you already have
- Good for systems with modest VRAM
GGUF route
- Fastest generation
- Lower VRAM than full-precision models
- Works well with optimization nodes like stage attention and torch.compile
FP8 safetensors
- Heavier at ~13 GB
- Good option if you prefer direct non-quantized models

Choose based on your hardware constraints and how often you plan to switch models.

Summary of Findings

Speed
- AccVideo delivered up to 8.5x faster generation at 720p compared to standard Hunyuan Video.
- 5–15 steps was enough for clean output in most prompts.
- Typical short clips finished in 1–3 minutes.
Quality
- Comparable to Hunyuan Video’s baseline results.
- Reward MPS LoRA further improved consistency and stability.
Flexibility
- Two clean routes in ComfyUI: LoRA-over-base and direct GGUF/FP8.
- GGUF Q8 is the quickest for iteration; LoRA stacking gives you fine-grained control.
Practicality
- Easy to add to existing Hunyuan workflows.
- Light on storage and VRAM when using LoRA or GGUF.

Final Notes

I’m currently using 10 steps for most runs and keeping frame counts around 90–120 at 720p. That setting balances speed and stability well, and it’s fast enough to iterate through prompts without long waits.

If you already work with Hunyuan Video, AccVideo is worth setting up. The speed improvement is real, and the integration in ComfyUI through either the LoRA route or the GGUF loader makes it simple to adopt.