Wan 2.2 Lightx2v LoRA: Versions, Settings and Slow Motion

Quick answer

The Wan 2.2 lightx2v LoRA is a pair of files, one for the high-noise expert and one for the low-noise expert of the Wan 2.2 14B models, that cuts sampling to 4 steps at CFG 1. lightx2v publishes them as Wan2.2-Lightning. They exist only for the 14B T2V and I2V models; the 5B is still an open item on lightx2v's to-do list.

What most people need:

• The files ComfyUI uses are the T2V V1.1 pair and the I2V V1 pair. Comfy-Org renamed them wan2.2_t2v_lightx2v_4steps_lora_v1.1_{high,low}_noise.safetensors and wan2.2_i2v_lightx2v_4steps_lora_v1_{high,low}_noise.safetensors. They are byte-identical to lightx2v's originals, at 1.23 GB each.
• The settings in both lightx2v's own workflows and ComfyUI's templates: 4 steps in total, the high-noise expert on steps 0 to 2 and the low-noise expert on steps 2 to 4, CFG 1, euler with the simple scheduler, shift 5, LoRA strength 1.0 on each expert.
• The templates ship with the LoRA off. The current Wan 2.2 14B T2V and I2V templates have an on/off switch that defaults to the plain 20-step path at CFG 3.5. Turn it on to get the 4-step settings above.
• Slow motion is a known, acknowledged problem. lightx2v said in August 2025 that its 4-step videos move more slowly than the base model's. Its later T2V releases, 250928 and V2.0, were aimed at that. No newer I2V release is in the Lightning repository.
• "V2" means T2V V2.0, released in November 2025. There is no I2V V2.0.

We have not run Wan 2.2 or these LoRAs ourselves. File sizes and dates below were read from the Hugging Face API and commit history on 2026-10-09; settings were read from the workflow JSON files themselves.

Every Wan2.2-Lightning release

All in lightx2v/Wan2.2-Lightning on Hugging Face. Dates are the first upload in the repository's commit history, in UTC. The rank was read from each file's tensor shapes. The weights are stored in FP32, which is why a rank-64 pair is larger than you might expect.

Inside each folder the files are simply high_noise_model.safetensors and low_noise_model.safetensors. lightx2v's own workflows load them by their folder path inside loras/. If you copy them loose into loras/ instead, rename them first, or one release overwrites the other.

What the README says about each:

• V1.1 is "slightly better than V1", in the README's words. V1 is still in the repository.
• 250928 is called a preview of V2.0, made with a new distillation method, with better camera control and motion. The release was dated rather than numbered, X-niper of lightx2v explained in the discussion thread, because there was still a quality gap to the base model.
• 250928 dyno replaces the high-noise expert with a fully trained model rather than a LoRA, and uses the 250928 low-noise LoRA. X-niper recommended it for highly dynamic scenes and warned it can produce blur, ghosting and intersecting geometry, especially at 720p. It needs its own workflow, published in the same folder.
• V2.0, announced on the GitHub README for November 8, 2025, claims better camera control and motion, and comes with the Phased DMD report (arXiv 2510.27684). The folder holds the two LoRAs and an empty config.json. No V2.0 workflow file was published.

A second set: Wan2.2-Distill-Loras

lightx2v/Wan2.2-Distill-Loras holds a different kind of LoRA. lightx2v said in a December 2025 reply that these were extracted with SVD from fully distilled models, while the Lightning ones were trained as LoRAs.

The 1022 I2V pair, and the LoRAs in lightx2v/Wan2.2-I2V-A14B-Moe-Distill-Lightx2v, are the only I2V LoRAs newer than I2V V1. Both are SVD extractions, according to lightx2v. The 1022 sample workflow is built for Kijai's WanVideoWrapper nodes, not native ComfyUI.

Which files the ComfyUI templates use

Comfy-Org added four lightx2v files to Comfy-Org/Wan_2.2_ComfyUI_Repackaged on 2025-08-11 and wired them into the templates the same day. We compared SHA-256 hashes: each is the same file as lightx2v's original.

The repack has no 250928, dyno or V2.0 files. To try those you download them from lightx2v and point the LoRA loaders at them yourself.

Keep T2V LoRAs on T2V workflows. X-niper warned that a T2V LoRA in an I2V workflow ruins the output.

The template notes list each LoRA as "1.14 GB" and each FP8 expert as "13.31 GB". Those are binary gigabytes (GiB) of the same byte counts. Hugging Face shows the decimal 1.23 GB and 14.29 GB. Which folder each file goes in is on our Wan 2.2 models folder page.

The settings: steps, CFG, shift and strength

Read from lightx2v's NativeComfy workflow files and from the Comfy-Org templates on main (last changed 2026-09-28).

Both use two KSampler (Advanced) nodes with the same seed. The first adds noise and returns with leftover noise; the second adds none. Shift is set with a ModelSamplingSD3 node on each expert. In the templates, a boolean named "Enable Lightning LoRA" (T2V) or "Enable 4steps LoRA?" (I2V) flips the model, steps, split point and CFG together, so you change one widget, not five.

High-noise expert + LoRA (strength 1.0) → ModelSamplingSD3 shift 5
  → KSampler (Advanced): steps 4, cfg 1, euler, simple, start 0, end 2,
    add_noise enable, return_with_leftover_noise enable
Low-noise expert + LoRA (strength 1.0) → ModelSamplingSD3 shift 5
  → KSampler (Advanced): steps 4, cfg 1, euler, simple, start 2, end 4,
    add_noise disable, return_with_leftover_noise disable

Why these numbers: lightx2v's notes for T2V V1 say it was distilled with the euler scheduler at shift 5, on the four timesteps 1000, 937.5, 833.33 and 625, and ask for CFG 1 so each step is one model pass, not two. Our arithmetic: shift 5 applied to a plain 4-step schedule gives exactly those four timesteps, so the ComfyUI setup above reproduces the training schedule.

The FLF2V template does it differently. Its lightning path is a separate group, bypassed by default, with the same 4 steps, CFG 1 and shift 5; the default path there runs 20 steps at CFG 4 and shift 8.

The dyno model is a different setup

The 250928 dyno release is not a LoRA pair. Its workflow loads Wan2.2-T2V-A14B-4steps-250928-dyno-high-lightx2v.safetensors in Load Diffusion Model in place of the high-noise expert, with no LoRA on it. The low-noise side is the FP16 low-noise expert plus the 250928 low-noise LoRA at 1.0. Steps, split, CFG, sampler and shift are the same as above. The dyno file is 28.58 GB in BF16, so it goes in diffusion_models/, and it is twice the size of the template's FP8 expert. That changes the VRAM picture; see our Wan 2.2 VRAM requirements page.

The slow-motion problem

Users reported it within days. A discussion opened on 2025-08-08 says character movement came out in slow motion with the LoRA compared with the base model. It drew more than 60 replies.

What lightx2v has said, in order:

• 2025-08-08. X-niper replied that the T2V LoRA slows motion down and they were working on it. The I2V LoRA was better, X-niper added, but still behind the 40-step base model.
• 2025-08-13. lightx2v opened its own thread confirming that 4-step videos move more slowly than the base model's, and pointed to community workflows that use more than 4 steps to work around it.
• 2025-09-28. With the 250928 release, X-niper gave in-house "motion dynamics" scores: 10.66 for the base T2V model, 7.76 for 250928 and 5.27 for V1.1. Higher means more motion. These are lightx2v's figures on its own test set; the method is not published in the thread.
• 2025-10-30. On GitHub, a user said the newer SVD-extracted I2V LoRA was still mostly slow motion. X-niper replied that these SVD versions have some artifacts and slower motion than the full distilled model, and that a directly LoRA-distilled I2V version would follow. As of 2026-10-09 the Lightning repository has no I2V release after V1.

The workaround users posted most is a three-sampler setup: one or more high-noise steps with no LoRA at CFG 3.5, then the high-noise expert with the LoRA at CFG 1, then the low-noise expert with the LoRA. The step counts and LoRA strengths people used vary from post to post, and each one is a single user's report. Some said it fixed the speed. Others said it overcooked other LoRAs or looked too contrasted, and it costs extra steps. One user also reported that slow motion got worse at lower resolutions, with 832×480 the slowest of three sizes tried. That is one person's test, not a confirmed cause.

So if 4-step T2V output looks slow, the documented options are: try 250928 or V2.0, or the dyno model for fast scenes, or give up some speed with a no-LoRA first step. For I2V nobody has published a fix that lightx2v endorses.

Can you use the Wan 2.1 lightx2v LoRA on Wan 2.2?

People do, and it is a common source of "v2" confusion. The Wan 2.1 LoRAs that circulate, such as lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors (1,253,192,432 bytes, 1.25 GB, in Kijai/WanVideo_comfy), are made for Wan 2.1 14B. Their "v2" has nothing to do with Wan2.2-Lightning V2.0.

• In August 2025, a user reported that the Wan 2.1 lightx2v LoRAs gave better results on Wan 2.2 than Wan2.2-Lightning, while noting ComfyUI prints a loading warning.
• X-niper answered that lightx2v had compared Wan2.2-Lightning with the Wan 2.1 T2V weights on a test set before release, and found a large improvement.
• In October 2025, another user said their best results were still with the Wan 2.1 rank-128 T2V LoRA.

That is two user reports against lightx2v's unpublished test. Nobody has published a controlled comparison.

Speed: whose numbers

The same template notes put VRAM usage at 84% without the LoRA and 83 to 89% with it. In those notes the LoRA changes time far more than memory: the pair adds 2.45 GB on disk and cuts the steps.

We have not timed these LoRAs. The closest thing we have measured is a different model: MiniMax H3 has its own lightx2v 4-step LoRA, covered in our MiniMax H3 Turbo LoRA guide. Its files and settings do not carry over to Wan 2.2.

Common mistakes

• Using the templates as shipped and expecting 4 steps. The switch is off by default.
• Following the ComfyUI tutorial's lightx2v link. It points at Wan2.2-T2V-A14B-4steps-lora-rank64-V1, a folder lightx2v deleted on 2025-08-05. Use the repack or the -Seko- folders.
• Mixing T2V and I2V LoRAs, or putting the high-noise file on the low-noise expert.
• Raising CFG above 1 on the LoRA steps. lightx2v asks for CFG 1, and the templates set it. Higher values double the work per step; the three-sampler workaround only does it on a step without the LoRA.
• Black frames with a lightx2v file from outside the repack. One user saw this with a lightx2v FP8 file; see black video output in ComfyUI.

How to check a downloaded workflow's loaders, steps and links before you queue it is on our ComfyUI workflows page.

What nobody has published

• An official ComfyUI workflow or settings for T2V V2.0.
• An I2V LoRA after V1 trained the Lightning way, which lightx2v said would come.
• A 5B Lightning LoRA.
• The method behind lightx2v's motion dynamics scores, or the same scores for V2.0 or I2V.
• A measured comparison of the Wan 2.1 and Wan 2.2 lightx2v LoRAs on Wan 2.2.

Licence and downloads

The Wan2.2-Lightning README states the Apache 2.0 licence and adds use restrictions on illegal and harmful content; read it before building on the files. Comfy-Org/Wan_2.2_ComfyUI_Repackaged and lightx2v/Wan2.2-Distill-Loras carry the Apache 2.0 tag on Hugging Face. Kijai/WanVideo_comfy shows no licence tag; check the file's origin before you rely on it. The base Wan 2.2 models are Apache 2.0.

We do not host any of these files. Download them from the repositories named below.

To check a GPU and system RAM against a local video model, use the system requirements checker. Wan 2.2 is not one of its presets yet.

GenVidKit is an independent guide. It is not affiliated with Alibaba, the Wan team, lightx2v (ModelTC), Comfy Org, Kijai or Hugging Face.

Sources

All read on 2026-10-09.

• lightx2v/Wan2.2-Lightning on Hugging Face — file list, byte counts, commit dates, the README, the T2V V1 usage notes, and the NativeComfy workflow files.
• ModelTC/Wan2.2-Lightning on GitHub — release notes for 250928 and V2.0, the 5B to-do item.
• Wan2.2-Lightning discussions #20, #26, #14, #42, #43, #56 and #63 — the slow-motion reports and lightx2v's replies, the motion scores, dyno, the Wan 2.1 comparison, SVD versus trained LoRAs.
• GitHub issues #31 and #32 — the SVD-extracted I2V LoRAs and their slow motion.
• lightx2v/Wan2.2-Distill-Loras — the 1022 and 1217 files, sizes and dates.
• Comfy-Org/Wan_2.2_ComfyUI_Repackaged — repack file names, hashes, upload dates.
• Comfy-Org/workflow_templates — the switch, steps, CFG, shift and strength in video_wan2_2_14B_t2v.json, video_wan2_2_14B_i2v.json and video_wan2_2_14B_flf2v.json, and the RTX 4090D notes.
• Wan2.2 native workflow, ComfyUI documentation — the outdated lightx2v link.
• Kijai/WanVideo_comfy, Lightx2v/ — the Wan 2.1 LoRA file and size.
• Phased DMD preprint, arXiv 2510.27684 — the method lightx2v names for releases since September 28, 2025.