wanx-troopers
LoRa-s, Part I
LoRa-s Table Of Contents
- Loras Alchemy
- LoRa-s Part I; this document
- LoRa-s Part II
- Both Ways
2025.12.17
Purely experimental: add fastwan at 0.7 strength?…
2025.12.11
It is a common question which speed LoRA-s to use with Wan 2.2. For I2V models the advice is:
- on I2V high use either 1030 at strength 1 or Wan 2.1 480p at strength 3
- on I2V high use the same 480p either at a strength of either 1 or 1.5
1030 is generally good at prompt following but the human motion can be more basic and boring; 480p on high is more active but it has the tendency to activate the background as well and output can be erratic and inconsistent - but it can also be unexpectedly interesting as wel
Links to download both of the LoRA-s mentioned above are bellow on this page - in most of the cases to bf16 versions.
for 4 steps the scheduler matters a lot; lcm-beta? in the wrapper there’s the
flowmatch_distill
at least one step with cfg makes such a difference
“with cfg” means with cfg greater than 1, cfg of 1 disables it; generation step with cfg greater than 1 takes 2x time compared to a step with 1 (“without cfg”), therefore often just the very 1st step is done with cfg > 1
2025.11.23
FL13 three sampler setup for Wan 2.2 I2V
…or lightx2v_I2V_14B_480p_cfg_step_distill_rank64_bf16 on low
2025.11.13
Q: Best Wan 2.2 I2V LoRa-s? Artist: 2.1 LightX2V for 2.2 low noise, new 2.2 LightX2V 4-step distill (1030?) for high noise? M: 1030 feels like the only alternative so far for the old 2.1 at higher strength (and it is high-noise only)
2025.10.14
There’s
- 2.1 Lightx2v, that works fine with 2.2 low noise
- 2.2 Lightning for both high and low (and it’s different versions), has exposure issue and not best motion
- 2.2 Lightx2v 4 step distill for high noise, that has 3 versions so far:
- the initial called “MoE”
- 1022 date update
- 1030 date update
2025.10.14 (this refers to high noise wan 2.2 too):
well the thing is that nothing beats the old 2.1 lightx2v still
6 steps, split in middle and first step with cfg 2.0
strength 3.0 for high and 1.0 for low
Old is possibly “the rank 64 one”
2025.10.18
There are only 3 options for the low noise:
- original 2.1
- lightx2v
- 2.2 Lightning rCM
If we don’t count accvid/fast wan etc. That, in my opinion, changes the output too much from original. Though with that criteria I’d forget the 2.2 Lightning as well pretty much
2025.10.19
Q: which one are you using still the i2v 480 or t2v rank 64 and 3.0 with high and 1 with low?..
A: Something like that
Q: and a pinch of cfg on high, right?
A: First step at least
2025-?-?
Users chatting:
Q: what the best lora for t2v ? lightx?
A: lightx quantile 0.15 Lora from Kijai, its very good; for high 3.0 low 1.0 Strength
2.2 T2V
| Repo | Lora | Generation | Approx. Released | Comment | |
|---|---|---|---|---|---|
| HF:lightx2v/Wan2.2-Distill-Loras | wan2.2_t2v_A14b_high_noise_lora_rank64_lightx2v_4step_1217 wan2.2_t2v_A14b_low_noise_lora_rank64_lightx2v_4step_1217 |
2.2 T2V High/Low | 2025.12.17 | recommended? | |
| HF:Kijai/WanVideo_comfy_fp8_scaled/T2V lightxv2/Wan2.2-Lightning/dyno |
Wan2.2-T2V-A14B-4steps-250928-dyno-high-lightx2v model not LoRA | 2.2 T2V High model |
2025.09 | good motion | |
| HF:lightx2v/Wan2.2-Distill-Loras | wan2.2_t2v_A14b_high_noise_lora_rank64_lightx2v_4step_1217 wan2.2_t2v_A14b_low_noise_lora_rank64_lightx2v_4step_1217 |
2.2 T2V High/Low | 2025.12.17 | bf16 “not much better” |
|
| HF:lightx2v/Wan2.2-Lightning:Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V2.0 | Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V2.0 | 2.2 T2V High/Low | 2025.11.07 | No Kijai conversion available Reprotedly can make things too bright |
|
| HF:lightx2v/Wan2.2-Lightning:Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V1.1 | Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V1.1 | 2.2 T2V High/Low | 2025.08 ? | No Kijai conversion available | |
| HF:lightx2v/Wan2.2-Lightning:Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V1 | Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V1 | 2.2 T2V High/Low | 2025.08 ? | No Kijai conversion available | |
| HF:Kijai/WanVideo_comfy:LoRAs/Wan22-Lightning | Wan22_A14B_T2V_HIGH_Lightning_4steps_lora_250928_rank128_fp16 | 2.2 T2V High | - | - | |
| HF:Kijai/WanVideo_comfy:LoRAs/Wan22-Lightning | Wan22_A14B_T2V_LOW_Lightning_4steps_lora_250928_rank64_fp16 | 2.2 T2V Low | - | - | |
| HF:Kijai/WanVideo_comfy:LoRAs/rCM | Wan_2_1_T2V_14B_rCM_lora_average_rank_148_bf16 | 2.1 T2V | - | by Nvidia, ppl stopped talking about it seemingly | |
| HF:Kijai/WanVideo_comfy:Lightx2v | lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16 lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank256_bf16 |
2.1 T2V | - | Could be good on 2.2 T2V Low? | |
| HF:Kijai/WanVideo_comfy:Lightx2v | lightx2v_14B_T2V_cfg_step_distill_lora_adaptive_rank_quantile_0.15_bf16 | 2.1 T2V | - | Some artists find this good on 2.2 T2V Low, FL13 also recommends on 2.2 I2V Low |
2.2 I2V
| Repo | Lora | Generation | Approx. Released | Comment |
|---|---|---|---|---|
| HF:Kijai/WanVideo_comfy:LoRAs/Wan22_Lightx2v | recommended Wan_2_2_I2V_A14B_HIGH_lightx2v_4step_lora_v1030_rank_64_bf16 |
2.2 I2V High | 2025.10.30 | new and reportedly much better with prompt adherence Artist: “to be honest I stopped using this lora as I’m getting color and light shifts with it” remains the most recommended I2V high LoRa |
| HF:lightx2v/Wan2.2-Distill-Loras | wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_1022 | 2.2 I2V High | 2025.10.22 | superceeded by 1030 |
| HF:lightx2v/Wan2.2-Distill-Loras | wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_1022 | 2.2 I2V Low | 2025.10.22 | - |
| HF:Kijai/WanVideo_comfy:LoRAs/Wan22_Lightx2v | Wan_2_2_I2V_A14B_HIGH_lightx2v_MoE_distill_lora_rank_64_bf16 | 2.2 I2V High | - | superceeded by 1022 and 1030 ghosting with simple scheduler but not linear quadratic? can do good motion and camera motion |
| HF:Kijai/WanVideo_comfy:Lightx2v | recommended lightx2v_I2V_14B_480p_cfg_step_distill_rank64_bf16 lightx2v_I2V_14B_480p_cfg_step_distill_rank256_bf16 |
2.1 I2V | - | Recommended on 2.2 I2V Low with a strength of 1.0-1.5 could even be used on high with strength 3 |
| HF:Kijai/WanVideo_comfy:Lightx2v | lightx2v_14B_T2V_cfg_step_distill_lora_adaptive_rank_quantile_0.15_bf16 | 2.1 T2V | - | Some artists find this good on 2.2 T2V Low, FL13 also recommends on 2.2 I2V Low |
Demoted due to developers’ comments
The Wan2.2 rCM checkpoints are obtained by merging Wan 2.1 rCM weights to Wan 2.2 checkpoints, no extra training inlcuded. Should have the same effect as directly using the Wan 2.1 rCM LoRAs.
| Repo | Lora | Generation | Approx. Released | Comment |
|---|---|---|---|---|
| HF:Kijai/WanVideo_comfy:LoRAs/rCM | Wan22-I2V-A14B-HIGH-rCM6_0_lora_rank_64_bf16 | 2.2 I2V High/Low | 2025.12.09 | From NVidia |
Note: errors about img_emb keys in the console are expected if using lghtx2v 2.1 LoRA-s.
Can be ignored, just happens because 2.2 doesn’t have those layers.
2.1
| Repo | Lora | Generation | Approx. Released | Comment |
|---|---|---|---|---|
| Kijai/WanVideo_comfy | FastWan/ FastWan_T2V_14B_480p_lora_rank_128_bf16 |
2.1 T2V | - | very old but worth a try |
| vrgamedevgirl84/Wan14BT2VFusioniX | FusionX_LoRa/ Wan2.1_T2V_14B_FusionX_LoRA |
2.1 T2V | - | very good for 2.1 |
| vrgamedevgirl84/Wan14BT2VFusioniX | FusionX_LoRa/ Wan2.1_I2V_14B_FusionX_LoRA |
2.1 I2V | - | very good for 2.1 |
| vrgamedevgirl84/Wan14BT2VFusioniX | FusionX_LoRa/ Phantom_Wan_14B_FusionX_LoRA |
2.1 T2V | - | real trooper for 2.1 |
“Old” Lightning 2.2
HF:/Kijai/WanVideo_comfy:LoRAs/Wan22-Lightning/old contains fp16 versions of “old” Lightning LoRa-s. Seemingly these have disappeared from official “Lightning” repo but are preserved here. Both I2V and T2V high/low .safetensorts available, T2V in two versions - normal and v1.1:
Wan2.2-Lightning_I2V-A14B-4steps-lora_HIGH_fp16
Wan2.2-Lightning_I2V-A14B-4steps-lora_LOW_fp16
Wan2.2-Lightning_T2V-A14B-4steps-lora_HIGH_fp16
Wan2.2-Lightning_T2V-A14B-4steps-lora_LOW_fp16
Wan2.2-Lightning_T2V-v1.1-A14B-4steps-lora_HIGH_fp16
Wan2.2-Lightning_T2V-v1.1-A14B-4steps-lora_LOW_fp16
Of these Wan2.2-Lightning_I2V-A14B-4steps-lora_LOW_fp16 was recently successfully used with WanAnimate.
Non-obvious Combinations
The table bellow lists some non-obvious combinations of LoRa-s and models
| LoRAs | Models | Notes | HF fp8_e4m3 | HF bf16/fp16 |
|---|---|---|---|---|
| high:none low:Wan22_A14B_T2V_LOW_Lightning_4steps_lora_250928_rank64_fp16 |
high:Wan2.2-T2V-A14B-4steps-250928-dyno-high-lightx2v low:some wan 2.2 |
Good motion | high:Kijai/WanVideo_comfy_fp8_scaled/T2V | high:lightxv2/Wan2.2-Lightning/dyno low:Kijai/WanVideo_comfy/LoRAs/Wan22-Lightning |
| both high/low: lightx2v_14B_T2V_cfg_step_distill_lora_adaptive_rank_quantile_0.15_bf16.safetensors | Wan 2.2 T2V High/Low | - | - | Kijai/WanVideo_comfy |
| high: lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank256_bf16 low: ? |
Wan 2.2 T2V High/Low | 2.5 str on high | - | Kijai/WanVideo_comfy |
| high: lightx2v_I2V_14B_480p_cfg_step_distill_rank256_bf16 low: unknown |
high: new distill 4 step I2V base model low:regular Wan 2.2 I2V Low? |
accidentally combining seems to massively amp up the amount of motion in shots that just the distill or just the lora doesnt achieve on their own |
Note: LoRa extracted from “dyno” model looses important qualities of “dyno”; e.g. difference between “dyno” and base model cannot be captured by a LoRa.
Special Use
Wan 2.1 and 2.2 LoRA-s:
- HF:Ashmotv/water_morping_wan
- HF:Ashmotv/exploded_effect_wan exploded effect WAN; there should be a slow motion explode somewhere nearby
- HF:benjamin-paine/steamboat-willie-14b steamboat willy / steampunk willy
- Wan 2.2 Smartphone LoRa
- Cseti/wan2.2-14B-Kinestasis_concept-lora-v1
- Cseti/LTXV-13B-LoRA-Wallace_and_Gromit-v1 aka Walgro
- Charlietooth/wan2.1-14B-lora-punk-style - a mix of cartoonish and realistic look
- Nebsh’s Lazy Susan Rotation Wan 2.2 trained on I2V but works on T2V too. (Simple but costly way to train: https://wavespeed.ai/models/wavespeed-ai/wan-2.2-i2v-lora-trainer, this one at 150steps was $7.5, trained on 6 videos from Higgsfield).
- Car/Clorth/Product consistency: drnighthan on HuggingFace
- Gouache - lora and workflows, for Wan 2.1 14B I2V
For morphing/AnimateDiff LoRA-s pls see
Wan 2.1 to Wan 2.2 LOW LoRA
From drozbay