Wan SVI Pro Advanced I2V

This node is a variant of the pack's main chaining node (WanAdvancedI2V), specifically tuned for SVI - Stable Video Infinity, EPFL's error-recycling technique that's the closest thing local video generation has to actually solving Wan's 5-second ceiling. If WanAdvancedI2V is the general-purpose chaining node, this is the one built for people running SVI's LoRAs and pushing resolution and length at the same time.

What SVI actually is, and why this node exists

SVI is a pair of LoRA adapters, not a model swap. Instead of just conditioning on a start image, SVI trains Wan on its own generation errors so that when you chain clips, the model learns to correct drift rather than compound it - that's the "error recycling" in the name. It's genuinely the biggest jump in open local video generation over the past year; the community thread announcing a 20-second continuous 1280×720 clip in 340 seconds is the highest-scoring Wan post in the whole corpus, and it was rendered through this exact pack.

Here's the thing worth knowing before you dive in: this pack is one of the primary places people actually run SVI from. The lab's own GitHub repo ships the training code and the paper; the versions people run day to day come from third parties - Kijai hosts the SVI v2 PRO LoRA weights, and wallen0322's ComfyUI-Wan22FMLF supplies the workflow and the node logic that makes them chain cleanly. This node is that logic, refined further for high resolution.

The README is specific about what problem this particular node fixes versus the plain SVI mode in WanAdvancedI2V: at high resolutions (1920×1080-class), SVI's motion transfer between chunks was measurably weaker, and stitched segments would occasionally skip frames at the boundary. This node is the author's rewrite to fix both, plus a simplified parameter set.

Inputs and outputs that matter

Required: positive, negative, vae, width, height, length, batch_size - plus two SVI-specific required parameters:

• motion_influence (default 1.0) - the weight given to motion carried over from the previous chunk. This is the core SVI dial: how much of what happened in the last clip's final frames should visibly continue into this one.
• overlap_frames (default 4, range 4-128, step 4) - how many frames of the previous chunk overlap with this one during stitching. Higher overlap gives the model more context to blend the seam, at the cost of more redundant compute per chunk.

Optional, and this is where the "Pro" tuning lives:

• motion_boost (0.5-3.0, default 1.0) - amplifies motion amplitude specifically. If your chained clips feel like they're moving in slow motion (a known interaction between speed LoRAs and Wan chaining), this is the first dial to reach for.
• detail_boost (0.5-4.0, default 1.0) - the README's own gloss for this one is about dynamic pacing rather than static sharpness, so despite the name, treat it as another motion-character control alongside motion_boost rather than a texture/detail slider - worth confirming empirically on your own footage before you lean on it heavily.
• start_image, middle_image, end_image with matching enable_start_frame/enable_middle_frame/enable_end_frame toggles - the same first/middle/last anchoring as the rest of the pack, layered on top of the SVI continuation.
• middle_frame_ratio, high_noise_start_strength, low_noise_start_strength, high_noise_mid_strength, low_noise_mid_strength, low_noise_end_strength - the familiar per-frame, per-pass strength dials from the rest of the pack.
• prev_latent (LATENT) and video_frame_offset (INT) - the continuation hooks, fed from WanAdvancedExtractLastFrames' output and the previous chunk's next_offset, exactly like WanAdvancedI2V.
• clip_vision_start_image / _middle_image / _end_image - optional per-frame CLIP Vision conditioning.

Outputs: positive_high, positive_low, negative, latent, plus the chaining trio - trim_latent, trim_image, next_offset - for wiring straight into the next loop iteration.

Installing it

ComfyUI Manager: search "ComfyUI-Wan22FMLF". Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/wallen0322/ComfyUI-Wan22FMLF

Restart. This node alone doesn't ship the SVI weights - you need the actual SVI LoRA pair (Kijai's WanVideo_comfy repo hosts the current SVI v2 PRO HIGH/LOW rank-128 LoRAs, roughly 1.1GB each) loaded into your model chain upstream, plus the usual Wan 2.2 I2V-A14B high/low checkpoints and VAE.

Common issues

Progressive blur the longer the chain goes. This is a documented, repeated community complaint about SVI generally, not specific to this node's tuning - the effect compounds with length and gets worse when other motion LoRAs are stacked on top. There's no clean fix beyond accepting shorter chains or periodically re-anchoring with a fresh, sharp start image.

ComfyUI crashing during assembly, not generation. The most consistent failure mode for long SVI chains isn't the sampler, it's stitching many chunks together - reported crashes and OOMs during the final assembly step on setups with as much as 16GB VRAM / 64GB system RAM once total length crosses roughly a minute. If you're chaining that far, plan to save intermediate chunks and stitch outside ComfyUI.

Facial identity drifting across chunks. SVI conditions for length, not for identity - it's explicitly a known limitation, not a bug you can dial away with motion_influence. If a consistent face matters, pair this with a character LoRA on the base model rather than expecting SVI alone to hold it.

Slow-motion artifacts. Inherited from LightX2V-style speed LoRAs stacked underneath, not from SVI itself, but it's what you'll experience as "SVI looks slow." Try motion_boost above 1.0 first, and if that's not enough, check whether a speed LoRA on your high-noise pass is the real culprit - the standing Wan 2.2 advice is to keep speed LoRAs off the high-noise pass entirely.