ERNIE-Image × Wan 2.6 Image-to-Video Complete Workflow: ComfyUI Pipeline from Static Images to 1080P Animation

Build a complete "text-to-image → image-to-video" automated pipeline in ComfyUI using ERNIE-Image for static images and Wan 2.6, LTX 2.3, and other cutting-edge video models.

From Stable Diffusion to FLUX.2, open-source image generation has gone through multiple paradigm shifts. But one persistent challenge remains: how do you bring static images to life?

As of June 2026, video generation models like Wan 2.6, LTX 2.3, and Seedance 2.0 have matured to support 1080P output. ERNIE-Image, as an 8B-parameter open-source text-to-image flagship, provides the ideal starting point for high-quality static images.

This article walks you through chaining these technologies into a complete ComfyUI workflow — turning AI-crafted static images into smooth dynamic videos.

Why ERNIE-Image as the Starting Point

ERNIE-Image brings three core advantages to image-to-video scenarios:

High-quality static images: The 8B DiT architecture ranks among the top open-source models for character consistency, scene detail, and lighting.

Precise instruction following: ERNIE-Image's 3B Prompt Enhancer and strong instruction-following capabilities ensure generated static images align closely with your creative vision.

Broad style coverage: From photorealistic photography to anime illustration, ERNIE-Image's style versatility provides a rich starting point for different types of video content.

2026 June Video Model Landscape

Model	Parameters	Key Strength	VRAM Needed	Best For
Wan 2.6	14B MoE	Motion quality, native 1080P	24GB+	General I2V, cinematic motion
LTX 2.3	2.6B	Fastest speed, quantization-friendly	8GB+	Quick iteration, low-end hardware
Seedance 2.0	Unknown	Multi-shot consistency, audio sync	24GB+	Multi-shot films, ads
Kling 3.0	Closed	4K output, character consistency	API	Character-driven narratives, long video

Wan 2.6 is the best open-source choice — its MoE architecture controls inference costs while maintaining motion quality, and ComfyUI official nodes provide native support.

Base Workflow: ERNIE-Image → Wan 2.6 I2V

Environment Setup

# 1. Install ComfyUI
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
pip install -r requirements.txt
2. Download ERNIE-Image model
From https://huggingface.co/Comfy-Org/ERNIE-Image
Place in models/diffusion_models/
3. Download Wan 2.6 model
From https://huggingface.co/Wan-AI
Place in models/unet/ or models/diffusion_models/

Workflow Node Chain

[CLIP Text Encode] → [ERNIE-Image Sampler] → [Save Image]
                                                        ↓
[Load Image] ← [Generated Static Image]
     ↓
[Wan2.6 ImageToVideo] → [Save Video]

ERNIE-Image Generation Config

{
  "model": "ernie-image-base.safetensors",
  "vae": "flux2_vae.safetensors",
  "clip": "clip_l.safetensors",
  "scheduler": "dpmpp_2s_ancestral",
  "steps": 8,
  "cfg": 5.0,
  "width": 1024,
  "height": 768,
  "use_pe": true
}

Key tip: For images destined for video conversion, use 1024×768 resolution (16:9) — this is both ERNIE-Image's sweet spot and Wan 2.6's optimal input size.

Wan 2.6 I2V Config

{
  "model": "wan2.6_i2v_fp8.safetensors",
  "vae": "wan_vae.safetensors",
  "motion_bucket_id": 128,
  "frames": 81,
  "fps": 24,
  "resolution": "1024x576",
  "guidance_scale": 6.0,
  "steps": 30
}

Parameter Tuning Guide:

• motion_bucket_id: Controls motion intensity (64=subtle, 128=medium, 256=strong)
• frames: 81 frames ≈ 3.4 seconds @24fps
• resolution: Wan 2.6 supports up to 1080P, but 1024×576 is most stable on 24GB VRAM

Advanced Workflow A: ERNIE-Image → Wan 2.6 → 1080P Upscale

Wan 2.6 natively supports 1080P output — no more need for Topaz Video AI post-processing:

[ERNIE-Image generates 1024x768]
     ↓
[Wan 2.6 I2V @1024x576, 81 frames]
     ↓
[WAN 2x Upscaler VAE] → 2560x1280 output
     ↓
[Save Video @1080P]

The WAN 2x Upscaler VAE node can upscale 1024×576 video to 2560×1280 (2.5K), maintaining motion consistency and visual quality.

Advanced Workflow B: Multi-Model Selection Guide

Low VRAM Setup (8-12GB)

[ERNIE-Image Turbo NVFP4 @4.78GB]
     ↓
[LTX 2.3 I2V @quantized]
     ↓
[Save Video @720P]

LTX 2.3 is currently the lightest I2V model, supporting FP8 quantization for 8GB VRAM operation. The downside is lower motion quality compared to Wan 2.6, but the speed advantage is significant.

Balanced Setup (16-24GB)

[ERNIE-Image FP8 @14GB]
     ↓
[Wan 2.6 I2V FP8 @20GB]
     ↓
[Save Video @1080P]

FP8 versions of both ERNIE-Image and Wan 2.6 can run sequentially on 24GB cards (RTX 4090, RTX 5070 Ti).

Flagship Setup (48GB+)

[ERNIE-Image BF16 @29GB]
     ↓
[Wan 2.6 FP16 I2V @48GB]
     ↓
[WAN 2x Upscaler VAE]
     ↓
[Save Video @4K]

Dual A100 or RTX 6000 Ada configurations can handle the full-precision pipeline.

Batch Production Pipeline

Combining with the multi-prompt batch generation workflow from EI-093, you can build a complete batch image-to-video pipeline:

[Prompt Multi Batch (100 different prompts)]
     ↓
[ERNIE-Image batch static image generation]
     ↓
[Wan 2.6 I2V batch video conversion]
     ↓
[Batch output: 100 short videos]

Real-world case: E-commerce product showcase video batch generation — use ERNIE-Image to create 100 product static images in different scenes, then convert them all to videos through Wan 2.6.

Hardware Requirements Summary

Setup	VRAM	Runnable Pipeline	Output Quality
RTX 4060 Ti 16GB	16GB	ERNIE-Image FP8 + LTX 2.3 quantized	720P
RTX 4070 Ti 16GB	16GB	ERNIE-Image FP8 + Wan 2.6 FP8 (reduced res)	720P-1080P
RTX 4090 24GB	24GB	ERNIE-Image FP8 + Wan 2.6 FP8	1080P
RTX 5090 32GB	32GB	ERNIE-Image FP8 + Wan 2.6 FP16	1080P-4K
A100 80GB	80GB	ERNIE-Image BF16 + Wan 2.6 FP16	4K

Common Troubleshooting

Issue 1: Wan 2.6 node can't find model file

• Check that the model filename matches exactly what the node README specifies
• Confirm file is placed in models/unet/ or models/diffusion_models/

Issue 2: Too few video frames

• Increase the frames parameter (81=3.4s, 161=6.7s)
• Lower resolution to free up VRAM

Issue 3: Unnatural motion

• Adjust motion_bucket_id (64=subtle, 128=medium)
• Ensure ERNIE-Image generates high-quality static images (motion quality depends on input quality)

Issue 4: Out of VRAM

• Use FP8 quantized versions
• Lower output resolution and frame count
• Add a Clear Cache node before Wan 2.6 to free ERNIE-Image's VRAM

Summary

The ERNIE-Image + Wan 2.6 combination gives the open-source community its first complete "text → static image → dynamic video" pipeline. The 8B-parameter ERNIE-Image provides high-quality static starting points, while the 14B MoE Wan 2.6 brings those images to life.

Compared to the LTX 2.3 / Wan 2.2 approach covered in EI-056, Wan 2.6 delivers three substantive upgrades: native 1080P support, smoother motion quality, and better ComfyUI official node integration. For content creators, this means a complete creative-to-production pipeline is now possible on a single 24GB consumer GPU.