═══════════════════════════════════════════════════════════════════════════════ FPV DRONE DIVE PERSPECTIVE - WAN 2.2 ═══════════════════════════════════════════════════════════════════════════════ ─────────────────────────────────────────────────────────────────────────────── IDENTITY ─────────────────────────────────────────────────────────────────────────────── Base model Wan 2.2 Generation mode not declared by the author Checkpoint not determined — verify before use Category Camera movement Author ErosX Source https://civitai.com/models/2191633 Provenance CivitAI · model 2191633 · version 2467713 ─────────────────────────────────────────────────────────────────────────────── WEIGHTS ─────────────────────────────────────────────────────────────────────────────── File FPV_drone_dive_high_noise.safetensors Version High Size 146.3 MB Format SafeTensor / unknown precision SHA256 35704a53938f37d9bfa268394464e0469673e7b88dcd74fbfd8401739605be47 Published unknown ─────────────────────────────────────────────────────────────────────────────── TRIGGER WORDS ─────────────────────────────────────────────────────────────────────────────── No trigger word. The LoRA applies as soon as it is loaded — strength is the only way to dial the effect up or down. ─────────────────────────────────────────────────────────────────────────────── RECOMMENDED SETTINGS ─────────────────────────────────────────────────────────────────────────────── The author did not publish specific settings. No author settings were published. Community baseline: steps Official Wan2.2 reference code (wan/configs/): sample_steps 40 for T2V-A14B and I2V-A14B with the expert handover at the SNR boundary, and 50 for TI2V-5B. ComfyUI's CURRENT 14B t2v/i2v templates ship only the distilled configuration: 4 steps total, split 2 on the high-noise expert then 2 on the low-noise expert (two KSamplerAdvanced nodes, start/end 0->2 and 2->4). An undistilled 20-step configuration split 10/10 shipped as a second, bypassed node group in an earlier revision of the t2v template and still ships as the ACTIVE default in the current video_wan2_2_14B_flf2v template (20 steps, split 0->10 and 10->end). TI2V-5B: 20 steps in ComfyUI's template, 50 in the reference code; single sampler, no split., cfg Official reference sample_guide_scale is a per-expert pair: (3.0 low- noise, 4.0 high-noise) for T2V-A14B and (3.5, 3.5) for I2V-A14B — the two experts are allowed different CFG. TI2V-5B: 5.0 in both the reference code and ComfyUI's template. Undistilled ComfyUI values vary by template revision: 3.5 in the older t2v template's bypassed undistilled group, 4.0 in the current flf2v template. With ANY 4-step distill LoRA: cfg 1.0 on both samplers (verified in the shipped t2v and i2v templates)., sampler euler + simple scheduler on both A14B experts, with shift set via ModelSamplingSD3: 5.0 in ComfyUI's current distilled 14B t2v/i2v templates, 8.0 in the undistilled groups (older t2v revision, and the current flf2v template). Official reference sample_shift differs by task: 12.0 for T2V-A14B, 5.0 for I2V-A14B, 5.0 for TI2V-5B. TI2V-5B in ComfyUI: uni_pc + simple, shift 8.0., resolution A14B: 480P and 720P — 1280x720 or 832x480 (and portrait). Dimensions should be multiples of 16 (Wan2.1 VAE stride 4x8x8 combined with 1x2x2 patching, both read from the reference configs); ComfyUI's templates ship a cheap 640x640. Length 81 frames at 16 fps (~5s), frames 4n+1. TI2V-5B: 720P, 1280x704 or 704x1280; multiples of 32, because the Wan2.2-VAE stride is 4x16x16 and the model card states an additional patchification layer brings total compression to 4x32x32 — which is why it is 704, not 720. Length 121 frames at 24 fps (~5s)., lora_strength Official Lightning/lightx2v 4-step LoRAs: 1.0 on both experts — verified in ComfyUI's shipped t2v and i2v templates, where both LoraLoaderModelOnly nodes carry strength 1.0. Character/style/motion LoRAs are commonly run at 0.6-1.0, and when stacked on top of a 4-step distill LoRA the usual community move is to pull the high-noise-side strength down (roughly 0.5-0.8); those subject-LoRA numbers are community practice, not published by the authors — see "unverified".. ─────────────────────────────────────────────────────────────────────────────── ADOPTION & POPULARITY ─────────────────────────────────────────────────────────────────────────────── No statistics available. ─────────────────────────────────────────────────────────────────────────────── COMMUNITY NOTES ─────────────────────────────────────────────────────────────────────────────── 2 comment(s) on the source listing; showing 1. [+13] yooooo I’m gonna need @GayLizardSpy to make a FPV drone shot flying into a gaping asshole Archive summary: FPV Drone Dive Perspective - Wan 2.2 is a LoRA designed for Wan Video- family checkpoints, published by [ErosX](/users/ErosX) on CivitAI in December 2025, now on version 2. It layers on top of any compatible Wan Video checkpoint. You load the checkpoint first, then apply this LoRA on top of it. Wan 2.2 uses a two-pass architecture — a high-noise model handles motion and scene composition, and a low-noise model refines detail. Both models need to be loaded together for a complete pipeline; a single pass on its own won't produce usable output. When using speed LoRAs (lightx2v, CausVid, Lightning) Source: https://civarchive.com/models/2191633?modelVersionId=2467713 Comments are other users' words, reproduced as written. ─────────────────────────────────────────────────────────────────────────────── CHAINING / STACKING ─────────────────────────────────────────────────────────────────────────────── Stacks with others not stated The author gave no chaining guidance. The rules below are the general Wan 2.2 ones, not anything specific to this LoRA. General Wan 2.2 rules that apply here (see REFERENCE.txt): • The high/low pair is not a stack. Every A14B LoRA that ships two files needs the high-noise file on the high-noise UNET chain and the low- noise file on the low-noise UNET chain — two LoraLoaderModelOnly nodes on two different models, never both on one model. • Never load a high-noise LoRA onto the low-noise expert or vice versa. It loads without a single warning because the two checkpoints are architecturally identical, and the damage shows up only as degraded output. Filename is the only guard. • A single-file A14B LoRA with no high/low marking (common for Wan2.1-era LoRAs and for LoRAs extracted from merged models) goes on BOTH experts at the same strength. Loading it only on the low-noise expert is the safer half-measure when it fights the composition. • Stack subject/style LoRAs after the accelerator LoRA in the chain: UNETLoader -> Lightning LoRA -> subject LoRA -> ModelSamplingSD3. Order within LoraLoaderModelOnly chains is additive so it rarely matters mathematically, but keeping the accelerator first matches every reference workflow. • Do not stack two step-distillation LoRAs on the same expert. One accelerator per expert. • When adding subject LoRAs on top of a 4-step distill LoRA, reduce the subject LoRA strength rather than the distill LoRA strength — the distill LoRA is what keeps 4 steps viable at cfg 1.0. • Keep the LoRA strength for the two experts independently tunable. Because the high-noise expert owns layout and the low-noise expert owns detail, the community routinely runs a style LoRA weaker on high-noise and full-strength on low-noise. • TI2V-5B LoRAs chain like any single-model LoRA and can never be mixed into an A14B chain, or the reverse — the two have entirely different tensor geometry. • The author never states which weight family this targets. Three distinct answers. The structural half of each is verified from the shipped config.json and safetensors index files; the behavioural severity is community-reported and flagged in "unverified". (1) A14B <-> TI2V-5B: INCOMPATIBLE. The 5B model is dim 3072 / 30 layers / 24 head ─────────────────────────────────────────────────────────────────────────────── PROS ─────────────────────────────────────────────────────────────────────────────── + Well established — 2,021 downloads. [source: CivitAI stats] + Positively received — 98 thumbs-up with no down-votes. [source: CivitAI stats] + Compact at 146 MB — cheap to stack with other LoRAs. [source: file size] ─────────────────────────────────────────────────────────────────────────────── CONS & CAVEATS ─────────────────────────────────────────────────────────────────────────────── - No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata] - Author never states which generation mode it was trained for; test against your own workflow before relying on it. [source: description] - Thin documentation — expect trial and error. [source: description length] ─────────────────────────────────────────────────────────────────────────────── COMPATIBILITY WARNING ─────────────────────────────────────────────────────────────────────────────── Three distinct answers. The structural half of each is verified from the shipped config.json and safetensors index files; the behavioural severity is community-reported and flagged in "unverified". (1) A14B <-> TI2V-5B: INCOMPATIBLE. The 5B model is dim 3072 / 30 layers / 24 heads / ffn 14336 against the A14B's dim 5120 / 40 layers / 40 heads / ffn 13824 (both configs fetched and compared). In ComfyUI native this is SILENT DEGRADATION, not a hard error: keys for blocks 30-39 hit logging.warning("lora key not loaded: {}") in comfy/lora.py, and keys for blocks 0-29 match by name but not by shape, so they hit the shape guard logging.warning("WARNING SHAPE MISMATCH {} WEIGHT NOT MERGED") in comfy/lora.py or the caught exception logging.error("ERROR ...") in comfy/weight_adapter/lora.py, which returns the weight unmodified. All three call sites were read in the current ComfyUI source. The queue completes, no exception reaches the UI, and the output is effectively that of the LoRA node bypassed. (2) T2V-A14B <-> I2V-A14B: STRUCTURALLY COMPATIBLE, behaviourally risky. Verified from the safetensors index files: both experts carry exactly 1095 tensors with identical key names, and the only tensor whose shape differs is patch_embedding.weight (in_dim 16 vs 36), which LoRA trainers normally exclude. So a T2V-A14B LoRA loads onto I2V-A14B and vice versa with no warnings. What you get is a real but off- target effect. Treat cross-loading as "loads and does something", never as "certified". (3) High-noise expert <-> low-noise expert (within one A14B family): the two checkpoints have identical architecture, so a LoRA trained on one loads onto the other with zero warnings. This is the most dangerous case in the family, because there is no signal at all that you mismatched. Filename is the ONLY discriminator. (4) Wan2.1 14B LoRAs. Verified: Wan2.1-T2V-14B's config.json is dim 5120 / ffn 13824 / 40 layers / 40 heads / in_dim 16 — the same block geometry as Wan2.2-A14B, so Wan2.1 T2V LoRAs load. Wan2.1-I2V-14B is only a partial match: its index has 1303 tensors including 8 img_emb.* keys and 200 cross_attn k_img/v_img keys for its CLIP path, which Wan2.2-I2V dropped entirely (0 such keys); those will be skipped with "lora key not loaded". ─────────────────────────────────────────────────────────────────────────────── AUTHOR'S DESCRIPTION ─────────────────────────────────────────────────────────────────────────────── (the author left no description) ─────────────────────────────────────────────────────────────────────────────── EXAMPLE MEDIA ─────────────────────────────────────────────────────────────────────────────── 6 preview(s) in ./example_images/ — community results for this LoRA. Videos keep a matching _poster.jpg still frame. No prompts are recorded below: CivitAI's public API flags these posts as having generation metadata but does not return it, so the prompts cannot be scraped. Open the model page to read them. 01_video.mp4 (video, 1088x832) still: 01_poster.jpg 02_video.mp4 (video, 720x480) still: 02_poster.jpg 03_video.mp4 (video, 720x480) still: 03_poster.jpg 04_video.mp4 (video, 720x480) still: 04_poster.jpg 05_video.mp4 (video, 720x480) still: 05_poster.jpg 06_video.mp4 (video, 720x480) still: 06_poster.jpg ─────────────────────────────────────────────────────────────────────────────── HOW TO USE (COMFYUI) ─────────────────────────────────────────────────────────────────────────────── 1. Put every LoRA .safetensors in ComfyUI/models/loras/ (subfolders are fine and are used by reference workflows). Confirmed by the file-layout tree inside ComfyUI's own video_wan2_2_14B_t2v template note. 2. Install the base weights alongside them. A14B (from Comfy- Org/Wan_2.2_ComfyUI_Repackaged, split_files/): diffusion_models/wan2.2_{t2v,i2v}_high_noise_14B_fp8_scaled.safetensors AND the matching _low_noise_ file — you need both; text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors; vae/wan_2.1_vae.safetensors. TI2V-5B: diffusion_models/wan2.2_ti2v_5B_fp16.safetensors, the same umt5 text encoder, and vae/wan2.2_vae.safetensors. 3. Load the base model with UNETLoader ("Load Diffusion Model"), NOT CheckpointLoaderSimple — Wan2.2 ships as a bare diffusion model, with the text encoder loaded separately by CLIPLoader (type "wan") and the VAE by VAELoader. This is the node set used in every shipped Wan2.2 template. 4. For A14B, build two parallel chains, one per expert: UNETLoader (high_noise) -> LoraLoaderModelOnly -> ModelSamplingSD3(shift) -> KSamplerAdvanced, and separately UNETLoader (low_noise) -> LoraLoaderModelOnly -> ModelSamplingSD3(shift) -> KSamplerAdvanced. Use LoraLoaderModelOnly, not LoraLoader — Wan LoRAs patch the DiT only, there is no CLIP side to patch. 5. Put the high-noise LoRA file on the high-noise chain and the low-noise LoRA file on the low-noise chain. If a LoRA ships as a single file with no high/low marking, loading the same file on both chains is the standard community handling. 6. Wire the two KSamplerAdvanced nodes in series on one latent: the first (high-noise) runs add_noise=enable, start_at_step 0, end_at_step = boundary, return_with_leftover_noise=enable; the second (low-noise) runs add_noise=disable, start_at_step = boundary, end_at_step = total, return_with_leftover_noise=disable. Boundary is step 2 of 4 with a 4-step Lightning LoRA, step 10 of 20 undistilled. 7. For TI2V-5B there is only one chain and one KSampler: UNETLoader -> LoraLoaderModelOnly -> ModelSamplingSD3 -> KSampler. No expert split, no second sampler. 8. Fastest correct start: load ComfyUI's built-in templates video_wan2_2_14B_t2v / video_wan2_2_14B_i2v / video_wan2_2_14B_flf2v / video_wan2_2_5B_ti2v (Workflow > Browse Templates) and swap the LoRA filenames — the 14B templates already ship the two-expert plumbing and the Lightning LoRA nodes wired correctly. Note the 14B t2v/i2v templates are now packaged as a subgraph. ─────────────────────────────────────────────────────────────────────────────── Scraped by lorakit on 2026-09-05 03:49 UTC. Stats and description are the author's; pros/cons are derived from the evidence tagged beside each line. ───────────────────────────────────────────────────────────────────────────────