These are mixtures of WAN 2.2 and other WAN-like models and accelerators (with CLIP and VAE also included) to provide a fast, "all in one" solution for making videos as easily and quickly as possible. FP8 precision. Generally the latest version available for each type of model (image to video or text to video) is recommended.
MEGA Merge: This is the "one model to rule them all" version which pretty much does everything. It can handle text to video, image to video, and first frame to last frame and last frame only (because it includes VACE). There is a specific workflow to use these merges included in the mega-v3/ folder, as it is slightly more complicated (but shouldn't be slower) due to its flexibility. See below for a screenshot of "mega" being used.
NSFW Merges: Degenerates should steer clear of these merges, as they are only for the most civilized people of culture or scientific researchers. These merge various spicy WAN 2.1+2.2 LORAs at generally low strengths to provide a "jack of all trades, master of none" all in one despicable solution. If you are not getting the results you want, add more LORAs or just use the non-NSFW versions with hand-picked LORAs.
You just need to use the basic ComfyUI "Load Checkpoint" node with these, as you can take the VAE, CLIP and Model all from one AIO safetensors (saved in your 'checkpoints' folder). All models are intended to use 1 CFG and 4 steps. See sampler recommendations for each version below.
WAN 2.1 LORA compatibility is generally still good, along with "low noise" WAN 2.2 LORA compatibility (do not use "high noise" LORAs). You might need to adjust LORA strengths (up or down) to get results you want, though.
MEGA version workflow screenshot (you can use VideoCombine instead of Preview Image):
MEGA I2V: Just bypass the "end frame" so the "start frame" will be your I2V starting frame. Keep everything else the same.
MEGA T2V: Bypass "end frame", "start frame" and the "VACEFirstToLastFrame" node. Set strength to 0 for WanVaceToVideo.
MEGA Last Frame: Just bypass the "start frame" and keep "end frame". Keep everything else the same as in the picture.
MEGA First->Last Frame: Use it like shown in the picture above.
Older non-MEGA workflows (v10 and below):
Seems to work even on 8GB VRAM:
CHANGELOG/VERSIONS:
base: This is the first attempt and very "stable", but mostly WAN 2.1 with few WAN 2.2 features. sa_solver recommended.
V2: This is a more dynamic mixture with more WAN 2.2 features. sa_solver OR euler_a sampler recommended. Suffers from minor color shifts and noise in I2V, typically just at the start.
V3: This is a mixture of SkyReels and WAN 2.2, which should improve prompt adherence and quality. euler_a sampler recommended, beta scheduler. Suffers from minor color shifts and noise in I2V, typically just at the start.
V4: WAN 2.2 Lightning in the mix! euler_a/beta recommended. I2V noise and color shifting generally improved, but motion is a bit overexaggerated.
V5: Improved overexaggeration of I2V model. euler_a/beta recommended.
V6: New merging structure and overall significantly improved quality. I2V noise for the first 1-2 frames still exists, but it clears up much better than previous versions. Some WAN 2.1 LORAs at heavy strengths may cause up to 5 poor early frames with T2V, where discarding (or lowering strengths) may help. sa_solver/beta recommended. I2V rarely suffers from some dramatic scene shifts.
V7: I2V scene shifting should be fixed, but some I2V noise persists (generally for just the first 1-2 frames). No changes needed for the T2V model, so that remains at V6. sa_solver/beta recommended.
V8: T2V is now based entirely off of WAN 2.2 "low" (with PUSA, SkyReels and Lightning accelerators mixed in), which should resolve noise problems with it (8.1 adds more SkyReels). I2V scaled back some of the WAN 2.2 mix, which was contributing to noise problems. There still is some minor I2V noise, but more of a delicate balance of WAN 2.2 + SkyReels to keep decent motion and flexibility. Euler_a/beta recommended.
V9: Removed PUSA and SkyReels from the WAN 2.2-side of I2V (and completely from T2V). as I think PUSA/SkyReels wasn't consistently helping (and sometimes hurting) when applied to WAN 2.2. This should provide a more reliable base to work from. euler_a/beta recommended, but feel free to experiment with sa_solver/beta or others!
V10: Fixes wrong accelerators being used (now WAN 2.2 Lightning in I2V and an an adaptive rank Lightx2v along with WAN 2.2 lightning in T2V). I2V now has a tendency to zoom into whatever is going on in your prompt, which I believe comes from increased camera movement from Wan 2.2 Lightning and being less tied to your initial image as the video progresses (so, prompt accordingly). Euler_a/beta still seems good.
MEGA v1: This is likely how I will continue making models, as I don't need separate I2V and T2V versions. No noise problems with I2V anymore! MEGA v1 is based off of WAN 2.2 "low T2V", then adds VACE Fun, SkyReels, FunReward and the usual accelerator/CLIP/VAE mix. Use the included workflow. ipndm/sgm_uniform sampler/scheduler recommended.
MEGA v2: Removed the FunReward LORA, which was causing faces to shift. I did notice some minor face shifting in the NSFW merge remaining, which I think is due to the LORA mixture, but it has been improved. Also reduced some of the SkyReels LORA a bit. ipndm/beta recommended.
MEGA v3: Very different merging method using SkyReels 2.1 33% base and WAN 2.2 66% on top. I now also match accelerators for each version (2.1 and 2.2), then merge. I think this gets a better result by basing "mega" on models designed for 1 sampler (2.1) but then bringing in most of WAN 2.2 to lay on top. I'm hoping this will improve camera control, LORA compatibility and keeping facial features. ipndm/beta recommended.
Looking for GGUFs? Looks like DooFY87 on CivitAI has been doing that:
https://civitai.com/models/1855105/rapid-wan-22-i2v-gguf
Looking for FP16 precision? TekeshiX has been helping me build variants in FP16 format. These should be the V5 I2V model:
https://huggingface.co/TekeshiX/RAPID-AIO-FP16/tree/main
DISCLAIMER: As you may expect, some compromises had to be made to reach this level of speed and simplicity. If you want more complex workflows and longer generation times to run "full WAN 2.2"'s pair of models (which will give higher quality results), or control over accelerator LORAs included in this merge, there are many resources elsewhere to do that.
- Downloads last month
- -