18 KiB
Wan2.2 I2V-14B Diagnostic Report
This document records the systematic diagnostic methodology used to debug the Wan2.2 I2V-14B (Image-to-Video, 14 billion parameter) pipeline in mlx-video, along with every bug found, its root cause, and fix.
Table of Contents
- Overview
- Architecture Summary
- Diagnostic Methodology
- Bug 1: Text Embedding Cross-Contamination
- Bug 2: VAE Encoder Weights Excluded from Conversion
- Bug 3: RoPE Frequency Computation
- Bug 4: VAE Encoder Temporal Downsample Order
- Bug 5: Non-Chunked VAE Encoding
- Verified Correct Components
- Performance Optimizations
- Open Investigation: CFG Effectiveness
- Reference Implementation
- Useful Diagnostic Commands
Overview
The I2V-14B pipeline takes an input image and generates a video using a dual-model diffusion transformer. The initial implementation produced severely broken output — first frame showed the image, subsequent frames degraded to noise, checkerboard artifacts, or flat grey.
Through a systematic component-by-component comparison against the reference PyTorch implementation, five bugs were found and fixed. The approach was to verify each component in isolation numerically, then narrow down failures to the subsystem level.
Timeline of Symptoms
| Stage | Symptom | Root Cause |
|---|---|---|
| Initial | Grey/blurry frames after frame 1 | Non-chunked VAE encoding (Bug 5) |
| After chunked encoding fix | First frame OK, rest degrades to noise | Text embedding cross-contamination (Bug 1) + RoPE frequencies (Bug 3) |
| After text + RoPE fix | Severe 8px checkerboard on frames 4+ | VAE encoder temporal downsample order (Bug 4) |
| After VAE fix | Image in frames 0-3, grey frames 4+ | CFG effectiveness issue (open investigation) |
Architecture Summary
I2V-14B Pipeline:
Input Image → VAE Encoder → [16, T_lat, H_lat, W_lat]
↓
Mask Construction → [4, T_lat, H_lat, W_lat]
↓
y = concat(mask, encoded_video) → [20, T_lat, H_lat, W_lat]
↓
Noise [16, T_lat, H_lat, W_lat] + y → [36, T_lat, H_lat, W_lat]
↓
Dual DiT (40 layers, 5120 dim) × 40 denoising steps
↓
Denoised Latent [16, T_lat, H_lat, W_lat]
↓
VAE Decoder → Video [3, F, H, W]
Key parameters:
in_dim=36(16 noise + 4 mask + 16 image latents),out_dim=16- Dual model: HIGH noise (t ≥ 900) and LOW noise (t < 900)
- 40 steps, shift=5.0, guide_scale=(3.5, 3.5)
- Uses Wan2.1 VAE (z_dim=16, stride 4×8×8)
Diagnostic Methodology
1. Component-Level Numerical Verification
Each component was tested in isolation against the reference PyTorch implementation:
- Load identical inputs (same random seed, same image, same prompt)
- Run through reference (on CPU where possible) and save intermediate tensors as
.npy - Run through MLX with the same inputs
- Compare outputs with
np.abs(ours - ref).max()and relative difference metrics
Components tested this way:
- RoPE frequency parameters and rotation output
- Time embedding (sinusoidal → MLP → projection)
- Patchify (reshape+Linear vs Conv3d)
- Unpatchify (transpose-based vs einsum)
- Scheduler (UniPC) timesteps and step formulas
- VAE encoder output (frame-by-frame comparison)
- Text embeddings (per-model MLP output)
- Cross-attention K/V cache shapes
- Mask construction values
2. Artifact Analysis
When visual artifacts appeared, quantitative metrics were used to characterize them:
- Checkerboard metric: Difference between even-indexed and odd-indexed pixels at patch boundaries. Values > 20 indicate visible checkerboard.
- FFT frequency analysis: Power at the 8px spatial frequency (matches VAE stride). 3× normal power confirmed VAE-stride-aligned artifacts.
- Per-frame statistics: Mean, std, min, max for each decoded video frame to track temporal degradation.
- Frame difference:
mean(|frame[i] - frame[i-1]|)to measure motion vs static content.
3. Isolation Testing
- VAE round-trip test: Encode image+zeros → decode. If clean, VAE decoder is not the source.
- Single-step model output: Run one diffusion step and compare cond vs uncond predictions to check CFG effectiveness.
- Patchify/unpatchify synthetic test: Pass structured gradient through unpatchify to verify spatial ordering.
- Resolution sweeps: Test at 480×272, 640×384, 1280×720 to check resolution dependence.
- Step count sweeps: Test at 5, 20, 40 steps to distinguish convergence issues from model bugs.
4. Weight Comparison
Direct comparison of converted MLX weights against original PyTorch weights:
# Load both weight sets
pt_weights = torch.load("model.safetensors")
mlx_weights = mx.load("model.safetensors")
# Compare each key
for key in pt_weights:
diff = np.abs(np.array(pt_weights[key]) - np.array(mlx_weights[key])).max()
Expected: max diff ≈ 0.001 (bfloat16 rounding). Actual: confirmed for all keys.
Bug 1: Text Embedding Cross-Contamination
Symptom: Model ignores text prompt, generated frames lack semantic content.
Root Cause: For the dual-model architecture (high-noise and low-noise experts), text embeddings were computed using only low_noise_model.embed_text() and reused for both models' cross-attention K/V caches. The two models have different text embedding MLP weights — 42% relative mean difference in output.
How Found: Compared text_embedding_0.weight and text_embedding_1.weight between high_noise_model.safetensors and low_noise_model.safetensors. Found 17.9% and 26.3% relative differences in the weight matrices.
Fix: Compute separate text embeddings per model:
# Before (broken):
context_emb = low_noise_model.embed_text([context, context_null])
cross_kv = low_noise_model.prepare_cross_kv(context_emb) # used for BOTH models
# After (correct):
context_emb_low = low_noise_model.embed_text([context, context_null])
context_emb_high = high_noise_model.embed_text([context, context_null])
cross_kv_low = low_noise_model.prepare_cross_kv(context_emb_low)
cross_kv_high = high_noise_model.prepare_cross_kv(context_emb_high)
File: mlx_video/generate_wan.py (lines 333–349)
Commit: a85b1c21
Bug 2: VAE Encoder Weights Excluded from Conversion
Symptom: VAE encoder produces constant output regardless of input image (all-zero weights after conversion).
Root Cause: The conversion script only included encoder weights for model_type == "ti2v" (TI2V-5B), not for "i2v" (I2V-14B). Since load_vae_encoder() uses strict=False, missing encoder weights were silently ignored, resulting in random initialization.
How Found: Traced through convert_wan.py and found include_encoder = config.model_type == "ti2v". Cross-referenced with the fact that I2V-14B also requires a VAE encoder (for image conditioning).
Fix:
# Before:
include_encoder = config.model_type == "ti2v"
# After:
include_encoder = config.model_type in ("ti2v", "i2v")
Note: The user's specific model happened to be manually converted with encoder weights already present, so this fix was preventive for future conversions.
File: mlx_video/convert_wan.py (line 424)
Bug 3: RoPE Frequency Computation
Symptom: Progressive 2px checkerboard artifacts on generated frames, increasing with temporal distance from the conditioned frame.
Root Cause: The reference creates one frequency table via rope_params(1024, head_dim=128) producing 64 frequency exponents, which rope_apply then splits into temporal (22), height (21), and width (21) portions. This gives temporal axes LOW frequencies and spatial axes progressively HIGHER frequencies.
Our code called rope_params three times with different normalizations:
# WRONG: each axis gets full frequency range [0, 1)
freqs_t = rope_params(1024, d_t=44) # 22 exponents normalized by 44
freqs_h = rope_params(1024, d_h=42) # 21 exponents normalized by 42
freqs_w = rope_params(1024, d_w=42) # 21 exponents normalized by 42
The max frequency difference was ~1.0 (not a precision issue — a fundamental design bug). This affected all Wan models (T2V, I2V, TI2V).
How Found: Line-by-line comparison of rope_params usage between reference model.py (single call) and our model.py (three calls). Printed the actual frequency exponents to confirm the numerical divergence.
Fix:
# Single unified frequency table, split by rope_apply
self.freqs = rope_params(1024, dim // config.num_heads)
Impact: ~35% reduction in checkerboard metric, 55% reduction in FFT 8px-frequency power.
File: mlx_video/models/wan/model.py (lines 154–156)
Commit: 3da4a637
Bug 4: VAE Encoder Temporal Downsample Order
Symptom: Massive checkerboard artifacts aligned to VAE spatial stride (8px period). VAE encoder output for frames 1–4 showed decreasing std (0.37→1.19) while reference showed stable std (0.95→1.34).
Root Cause: The VAE encoder has 3 downsampling stages. Two perform spatial+temporal downsampling (downsample3d) and one performs spatial-only (downsample2d). The order matters:
Reference: [False, True, True] → stage 0: 2d, stage 1: 3d, stage 2: 3d
Ours: [True, True, False] → stage 0: 3d, stage 1: 3d, stage 2: 2d ← WRONG
This caused temporal downsampling to happen at the wrong resolution stages (96-dim instead of 384-dim), corrupting temporal feature propagation.
How Found: Installed einops in the reference environment and ran the reference PyTorch VAE encoder on CPU. Compared frame-by-frame latent output:
- Frame 0 matched exactly (diff=0.0000) — spatial-only processing was correct
- Frames 1–4 had massive differences — proved temporal processing was broken
Then traced through the reference _video_vae() function and found it sets temperal_downsample=[False, True, True], while our Encoder3d class used the wrong default [True, True, False].
Fix:
# In Encoder3d.__init__, change default:
temporal_downsample = [False, True, True] # was [True, True, False]
Impact: Encoder output now matches reference within float32 precision (max_diff=2.2e-5). Checkerboard metric dropped from 60–80 to 0.1–7.7.
File: mlx_video/models/wan/vae.py (line 370)
Commit: 3da4a637
Bug 5: Non-Chunked VAE Encoding
Symptom: First 4–5 frames grey, then blurred version of image appears.
Root Cause: The reference VAE encoder uses chunked encoding with temporal caching (feat_cache):
- Encode first frame alone (1 frame)
- Encode remaining frames in chunks of 4, with cached temporal features propagating across chunks
- Each
CausalConv3dcaches last 2 temporal frames from its output, prepending them to the next chunk's input
Our original implementation encoded all frames at once with zero-padded causal convolutions. The temporal feature propagation is fundamentally different because:
- Chunked: real features from previous chunks serve as causal context
- Non-chunked: zeros serve as causal context for the start
How Found: Studied the reference CausalConv3d caching mechanism (feat_cache, feat_idx) and traced the temporal dimension through all encoding stages. Confirmed that non-chunked encoding produces different output by comparing tensor shapes and values.
Fix: Implemented full chunked encoding with temporal caching:
- Added
cache_xparameter toCausalConv3d.__call__ - Added
feat_cache/feat_idxpropagation toResidualBlock,Resample,Encoder3d - Rewrote
WanVAE.encode()with chunked loop (1-frame first chunk, then 4-frame chunks) - 24 cache slots across the encoder (1 conv1 + 18 downsamples + 4 middle + 1 head)
File: mlx_video/models/wan/vae.py (multiple methods)
Commit: b6a94c4c
Verified Correct Components
These components were numerically verified against the reference and are not sources of bugs:
| Component | Method | Max Diff | Notes |
|---|---|---|---|
| Weight conversion | Direct tensor comparison | ~0.001 | bfloat16 rounding only |
| RoPE rotation | Standalone comparison (float32 vs float64) | 1.3e-5 | Complex vs real multiplication equivalent |
| Time embedding | Full MLP comparison (sinusoidal→embed→project) | 7e-4 | 0.03% relative |
| Patchify | Conv3d vs reshape+Linear | 3.5e-3 | 0.16% relative |
| Unpatchify | einsum vs transpose(6,0,3,1,4,2,5) | exact | Identical operation |
| Scheduler (UniPC) | Formula-level audit + timestep comparison | exact | Predictor, corrector, lambda, rhos all match |
| Mask construction | Value comparison | exact | [4, T_lat, H_lat, W_lat], first temporal=1 |
| CFG formula | Code audit | — | uncond + gs * (cond - uncond) correct order |
| VAE decoder | Round-trip test (encode→decode) | clean | No checkerboard in round-trip output |
| Cross-attention K/V | Shape and value audit | — | Batch dimension preserved correctly |
Performance Optimizations
Applied alongside bug fixes to improve inference speed:
Pre-Computation (Before Diffusion Loop)
- Cross-attention K/V caching: Precompute K/V projections for all 40 blocks once
- RoPE cos/sin precomputation: Build frequency tensors once instead of per-step broadcast/concat
- Attention mask precomputation: Build padding mask once, pass via kwargs
- Inverse frequency caching: Store sinusoidal
inv_freqin__init__instead of recomputing - Timestep list conversion:
sched.timesteps.tolist()before loop to avoid.item()sync
Per-Step Optimizations
- Single patchify + broadcast for CFG B=2: Detect identical batch inputs, patchify once and broadcast instead of duplicating the Linear projection
- Vectorized RoPE: When all batch elements share the same grid size, apply rotation to the full batch tensor instead of looping per element
- Redundant type cast removal: MLX type promotion handles
bfloat16 * float32 → float32automatically — removed 240 unnecessary graph nodes per step (6 casts × 40 blocks) - Euler scheduler sync fix: Pre-store sigmas as Python floats to avoid
.item()evaluation sync
TeaCache Integration
- Polynomial rescaling stays in MLX lazy graph (Horner's method)
- Single
.item()call on the accumulated distance for the skip/compute decision - Configurable threshold, retention steps, and cutoff steps
Open Investigation: CFG Effectiveness
Current symptom: After all bug fixes, generated video shows the input image in frames 0–3 (latent frame 0), then grey/flat frames for the rest.
Finding: A single forward pass at t=1000 shows cond and uncond predictions are nearly identical (|diff| mean = 0.01–0.035). With guide_scale=3.5, the CFG guidance term barely changes anything.
Possible causes under investigation:
- Cross-attention context flow — both cond and uncond may be receiving equivalent context
- The model may genuinely produce small cond/uncond differences for I2V (since both share the same y conditioning)
- The
embed_textmethod orprepare_cross_kvmay not properly separate B=2 batch elements - There may be an issue with how cross-attention K/V caches index into batch elements
Diagnostic approach: Compare cross-attention K/V cache values between cond (index 0) and uncond (index 1) to confirm they contain different embeddings.
Reference Implementation
The reference PyTorch implementation is at /Users/daniel/Projects/Wan2.2/:
| File | Contents |
|---|---|
wan/image2video.py |
I2V pipeline (y construction, mask, diffusion loop) |
wan/modules/model.py |
DiT model (forward pass, RoPE, patchify) |
wan/modules/vae2_1.py |
VAE encoder/decoder with chunked encoding |
wan/utils/fm_solvers_unipc.py |
UniPC scheduler |
wan/configs/wan_i2v_A14B.py |
Model configuration |
Key structural differences between reference and our implementation:
- Reference runs separate B=1 forward passes for cond/uncond; we batch as B=2
- Reference uses
torch.amp.autocast('cuda', dtype=bfloat16)with explicit float32 blocks; we cast via weight dtype - Reference uses
Conv3dfor patchify; we use equivalentreshape + Linear - Reference casts timesteps to
int64; we keep as float (diff < 1.0)
Useful Diagnostic Commands
Run I2V-14B generation
python -m mlx_video.generate_wan \
--prompt "A woman smiles at camera" \
--image start.png \
--model-dir /Volumes/SSD/Wan-AI/Wan2.2-I2V-A14B-MLX \
--num-frames 17 --steps 40 \
--height 384 --width 640 \
--output output_i2v.mp4
Check VAE encoder output
import mlx.core as mx, numpy as np
from mlx_video.models.wan.vae import WanVAE
# Load VAE and encode an image
latents = vae.encode(video_tensor) # [1, 16, T_lat, H_lat, W_lat]
for t in range(latents.shape[2]):
frame = np.array(latents[0, :, t])
print(f"Frame {t}: mean={frame.mean():.4f} std={frame.std():.4f}")
Analyze video frame quality
import cv2, numpy as np
cap = cv2.VideoCapture("output.mp4")
while True:
ret, frame = cap.read()
if not ret: break
# Checkerboard metric: high values indicate patch-boundary artifacts
checker = np.abs(frame[::2, ::2].astype(float) - frame[1::2, 1::2].astype(float)).mean()
print(f"std={frame.std():.1f} checker={checker:.1f}")
Compare weights between PyTorch and MLX
import torch, mlx.core as mx, numpy as np
pt = torch.load("model.pt", map_location="cpu")
mlx_w = mx.load("model.safetensors")
for key in sorted(pt.keys()):
if key in mlx_w:
diff = np.abs(pt[key].float().numpy() - np.array(mlx_w[key])).max()
if diff > 0.01:
print(f"LARGE DIFF {key}: {diff:.6f}")