How to Do Image to Video in ComfyUI: Wan 2.2, MiniMax H3 and LTX-2.5 (2026)

Turning a still picture into a moving clip is one of the most fun things you can do with ComfyUI. The hard part is choosing a model, because there are now four good ones and each needs different files.

This guide shows you how to do image to video in ComfyUI step by step, using the official templates. You get the right model for your GPU, every file and folder, the best settings and fixes for the errors beginners hit most.

How Image to Video Works in ComfyUI

An image to video (I2V) model takes your picture as the first frame and generates the frames that follow it. Your text prompt describes the motion: what moves, how the camera moves and what happens next.

In ComfyUI every I2V workflow has the same parts:

• Load Image: your starting picture.
• Text encoder: turns your prompt into something the model understands.
• Diffusion model: the video model itself, the biggest file.
• VAE: converts between pixels and the model’s compressed format (some models have a second VAE for audio).
• Sampler: runs the steps that build the video.
• Create Video / Save Video: joins frames into an MP4 at a set frame rate.

The official templates wire all of this for you. Several newer ones pack the nodes into a single subgraph, so you see one big node with a few settings; double-click it to see inside.

Which Image to Video Model to Use

These are the four I2V templates most ComfyUI users run in November 2026, from Comfy-Org’s official template list.

Download sizes come from the template index and include every file the template uses. VRAM is a separate question; see how much VRAM you need for AI.

Before You Start

• ComfyUI installed: see how to install ComfyUI if you have not yet.
• Up to date: templates need a minimum version, for example ComfyUI 0.30.0 for MiniMax H3, 0.32.0 for LTX-2.5 and 0.36.0 for FastVideo FastH3. Update first.
• Disk space: 20 to 60 GB free, depending on the model.
• System RAM: 32 GB minimum, 64 GB recommended for the 14B to 22B models, because ComfyUI moves models between RAM and the GPU.
• Optional: ComfyUI Manager to install missing nodes in one click.

Tip: on Windows portable, run update\update_comfyui.bat before opening a new template. Most “missing node” errors on video templates are just an old ComfyUI.

Your First Image to Video, Step by Step

These steps work for every model below. Only the template name changes.

1. Open the template browser. Click Workflow > Browse Templates, then the Video section. Type “image to video” in the search box.
2. Pick a template for your GPU from the table above and click it.
3. Download the models. ComfyUI shows a list of missing files with download buttons. Let it download them, or fetch them yourself into the folders listed in this guide.
4. Load your picture. In the Load Image (or Load First Frame) node, click “choose file to upload” and pick your image.
5. Write a motion prompt. Describe what moves, for example “She turns her head to the camera and smiles, hair moving in the wind, slow dolly in.”
6. Keep it short. Set duration to 2 to 5 seconds and keep the default resolution for the first run.
7. Press Run. The first run loads the models and takes longest. The video appears in the Save Video node and in ComfyUI/output/.

Once that works, change one thing at a time: prompt, then length, then resolution. Fix the seed so you can compare results.

Option 1: Wan 2.2 5B (8 to 12 GB)

Wan 2.2 5B is a “TI2V” model: the same file does text to video and image to video. It is the lightest official option and runs on mid-range cards.

Files come from Comfy-Org/Wan_2.2_ComfyUI_Repackaged (the text encoder from Comfy-Org/Wan_2.1_ComfyUI_repackaged). Template defaults:

• Size and length: 1280×704, 121 frames, 24 fps (about 5 seconds).
• Sampler: 20 steps, CFG 5, uni_pc, simple, with ModelSamplingSD3 shift 8.

To use your image, connect a Load Image node to the start_image input of the Wan22ImageToVideoLatent node. On 8 GB cards, drop the size to 832×480 or 640×352 and the length to 49 to 81 frames.

Note: the 5B model uses wan2.2_vae, while the 14B models use the older wan_2.1_vae. Mixing them up is a classic error.

Option 2: Wan 2.2 14B I2V With 4-Step LoRA

Wan 2.2 14B uses two experts: a high-noise model for the first steps (layout and motion) and a low-noise model for the last steps (detail). The template runs both in two KSamplerAdvanced nodes.

The template has two paths you can switch between:

Both use euler / simple, shift 5, 16 fps and 81 frames (about 5 seconds) at 640×640. The timings are Comfy’s own figures from the template note. For 16 GB setups and GGUF files, see my Wan 2.2 on 8 to 16 GB guide.

Option 3: MiniMax H3 (Video + Audio)

MiniMax H3’s “Image to Video” template is the most used I2V template in ComfyUI right now. It makes video with sound, and it can take a first frame, a last frame or both.

• Sampler: res_multistep, simple scheduler, 4 steps by default with a Turbo LoRA; 20 steps without one.
• Size: native short edge 768 px, up to 768×1344, multiples of 32.
• Length: set in seconds; the template snaps it to H3’s 17-frame blocks at 24 fps. The default is 2 seconds.
• First/last frame: connect only first_frame for normal I2V, or both to make the model animate between two pictures.

Want it faster? The FastVideo FastH3: Image to Video template uses an 8-step distilled checkpoint (fastvideo_fasth3_8step_v2_pruned_int8_convrot.safetensors) with the same encoder and VAEs. Full H3 setup, including 16 GB tips: MiniMax H3 ComfyUI guide and my best H3 image to video workflow.

Option 4: LTX-2.5 (Video + Audio, HD)

LTX-2.5 from Lightricks animates a first frame into HD video with synchronized audio. The template uses a distilled 22B model and a two-stage process: a small first pass, then a 2x latent upscale.

• Defaults: 1280×720, 5 seconds, 24 fps.
• Sampler: euler_ancestral with fixed ManualSigmas lists for each stage. Do not swap them for a normal scheduler.
• prompt_enhance: off by default. Turning it on expands short prompts and adds about 1 to 2 minutes.
• Auto Duration: LTX-2.5 can predict the clip length from the action in your prompt.

Note the latent_upscale_models folder: it is easy to drop the upscaler into upscale_models by mistake. For 16 GB and GGUF options, see my LTX-2.5 on 16 GB guide.

Settings That Matter Most

Warning: do not raise CFG on a speed-LoRA or distilled workflow. At CFG above 1 those models give burnt, flickering video.

How to Write Motion Prompts

The image already shows what things look like. Your prompt should mostly describe what changes.

• Subject motion: “she laughs and brushes her hair behind her ear”.
• Camera motion: “slow dolly in”, “camera orbits to the left”, “static camera”.
• Environment: “leaves blow across the street”, “steam rises from the cup”.
• Sound (H3 and LTX only): “soft rain, distant traffic, she says ‘good morning’”.
• Keep it to one or two actions for a 5-second clip.

Example that works well on all four models:

The cat on the windowsill slowly turns its head toward the camera and blinks. Rain streaks down the glass behind it, soft warm lamp light. Static camera, gentle motion, calm mood.

What Makes a Good Starting Image

• Match the aspect ratio to your video size, or crop it first. Stretched inputs give warped motion.
• Sharp and well lit: blurry or tiny images give mushy video. Upscale small pictures first; see how to upscale images with AI.
• Clear subject with space to move: a person cropped at the edge cannot turn without the model inventing the rest.
• Simple backgrounds keep motion stable; very busy scenes flicker more.

Need a starting picture? Make one locally with Qwen-Image or another image model, then animate it. If you would rather not install anything, my free photo to video AI tools roundup covers online options.

Troubleshooting

Red “missing node” boxes when opening the template

Update ComfyUI; video templates use new core nodes. If a box is still red after updating, use ComfyUI Manager > Install Missing Custom Nodes.

“Reconnecting” in the middle of a run

Usually system RAM ran out while loading a big model. Close other apps, raise the Windows page file, or use smaller files. Full fix: ComfyUI Reconnecting error.

CUDA out of memory

Lower the resolution first, then the length. Use the speed LoRAs and fp8 or GGUF files. More options in my CUDA out of memory fix.

The video does not look like my picture

Check the Load Image node is connected to the start image input. With Wan 5B, an unconnected start_image means you are doing text to video.

Video is too fast or slow

The frame rate does not match the model. Use 16 fps for Wan 2.2 14B and 24 fps for Wan 5B, MiniMax H3 and LTX-2.5.

Flickering, melting faces or burnt colours

Lower CFG (to 1 with speed LoRAs), shorten the clip and use a sharper starting image. Faces also hold up better at 768 px than at 512 px.

“Error: shape mismatch” or VAE errors

A file from another model is loaded. Each model needs its own VAE and text encoder; check against the tables above. Folder questions: ComfyUI model folders cheat sheet.

My Take

Image to video in ComfyUI is much easier than a year ago because the official templates download the right files for you. Start with the model that fits your card, make a short clip, and only then push length and resolution.

On a 16 GB card I reach for MiniMax H3 with the Turbo LoRA when I want sound, and Wan 2.2 14B with the 4-step LoRA when I want LoRA styles. On 8 to 12 GB, Wan 2.2 5B is the one that just works.

FAQ

What is the best image to video model for ComfyUI?

For most people on 16 GB or more, MiniMax H3 (with audio) or Wan 2.2 14B with the 4-step LoRA. On 8 to 12 GB, Wan 2.2 5B.

Can I do image to video in ComfyUI with 8 GB VRAM?

Yes. Use Wan 2.2 5B at 832×480 or smaller with 49 to 81 frames, or GGUF versions of Wan 2.2 14B with the 4-step LoRA.

How long does image to video take in ComfyUI?

Comfy measured about 71 to 97 seconds for a 5-second 640×640 Wan 2.2 14B clip with the 4-step LoRA on an RTX 4090D, and about 9 minutes without it.

Where do I put image to video models in ComfyUI?

Video models go in models/diffusion_models/, encoders in models/text_encoders/, VAEs in models/vae/ and speed LoRAs in models/loras/.

Which ComfyUI image to video models make sound?

MiniMax H3 and LTX-2.5 generate synchronized audio. Wan 2.2 makes silent video.

How do I make a longer video?

Generate several short clips and use the last frame of one as the first frame of the next, or use H3’s first/last-frame mode. Long single clips need much more VRAM.

Is ComfyUI image to video free?

Yes. ComfyUI and these models are free to run locally. Check each model’s licence before commercial use.

Sources

• Comfy-Org workflow templates: Wan 2.2 5B, Wan 2.2 14B I2V, MiniMax H3 I2V, FastH3 I2V and LTX-2.5 I2V JSON and notes
• Comfy-Org/Wan_2.2_ComfyUI_Repackaged
• Comfy-Org/MiniMax-H3 and lightx2v/Minimax-h3-Turbo
• Lightricks/LTX-2.5
• ComfyUI docs: updating ComfyUI