Existing H3 latent
24 × 57 × 48 × 84
Cross-VAE latent translation
H3 latent.
LTX-2.5 pixels.
A convolutional adapter directly translates MiniMax H3 video latents into the latent space decoded by the frozen LTX-2.5 Conv Video VAE—without running the LTX diffusion transformer, latent upsampler, or refiner.
Verified LTX-2.5 Conv VAE. The exact 1.45 GB VAE weights used for every target and decode have the same SHA-256 as the official LTX-2.5 Diffusers Conv Video VAE. The separate LTX-2.5 diffusion decoder has not been evaluated and is not represented here.
What runs
The actual decode path
Deployment starts from an existing H3 latent—no H3 encoder is needed. Source and teacher videos below are evaluation references only.
Lossless align
384 × 25 × 24 × 42
DeterministicConv adapter
194.76M parameters
LearnedLTX latent
128 × 25 × 24 × 42
LTX-2.5 Conv VAE
192 × 768 × 1344
Frozen28.7792PSNR ↑adapter decode vs LTX teacher
0.87486SSIM ↑adapter decode vs LTX teacher
28.7109PSNR ↑adapter decode vs H3 source
0.85911SSIM ↑adapter decode vs H3 source
34.9914PSNR ↑LTX teacher vs H3 source
0.94715SSIM ↑LTX teacher vs H3 source
0.2208dB to target ↓formal gap to 29.0 dB
194.76Mdeployed parameterscurrent formal Conv3D adapter
199.84Mprepared upper scaledense-temporal A/B · not promoted
Visual evidence
Source, teacher, prediction
Two preserved validation clips from the promoted checkpoint. The direct adapter decode is synchronized with both references.
H3 source · referenceLTX teacher · referenceDirect adapter decode
28.8195 dBPrediction vs teacher PSNR
0.89473Prediction vs teacher SSIM
0.02538Temporal gradient MAE ↓
Triptych assets are display encodes only. Metrics were computed from all 192 full-resolution decoded frames before web compression; SSIM follows the fixed deterministic frame-sampling protocol. “LTX-2.5” on this page always means the official Conv Video VAE, not the separate diffusion decoder.
Adapter design
Convolutional, not a Transformer
Geometry is handled deterministically, leaving the network to learn the nonlinear H3-to-LTX channel basis and decoder-visible residual.
194.76M factorized Conv3D adapter
- Input packing
- linear temporal anchor + 3 nearest temporal slots + 2×2 spatial pixel-unshuffle
- Residual trunk
- 22 blocks, width 752, expansion 2, GroupNorm with 16 groups
- Each block
- 1×3×3 spatial conv → depthwise 3×1×1 temporal conv → gated 1×1×1 channel MLP
- Skip path
- fitted affine 384→128 projection, frozen during nonlinear training
- Training precision
- BF16, activation checkpointing, gradient clip 1.0
Two-stage supervision
Latent pretraining:
e = ẑLTX − zLTX
MSE = mean(e²), L1 = mean(|e|)
Temporal = mean(|Δt ẑLTX − Δt zLTX|)
Spatial = 0.5·[mean(|Δh ẑLTX − Δh zLTX|) + mean(|Δw ẑLTX − Δw zLTX|)]
L = MSE + 0.25·L1 + 0.10·Temporal + 0.02·Spatial
Current native decoder-aware stage:
L = mean(e²) + 160·mean([clip(D(ẑLTX)) − clip(D(zLTX))]²)
D is the frozen LTX Conv VAE; pixel L1, SSIM, gradient, temporal, and perceptual weights are all 0.
e = ẑLTX − zLTX
MSE = mean(e²), L1 = mean(|e|)
Temporal = mean(|Δt ẑLTX − Δt zLTX|)
Spatial = 0.5·[mean(|Δh ẑLTX − Δh zLTX|) + mean(|Δw ẑLTX − Δw zLTX|)]
L = MSE + 0.25·L1 + 0.10·Temporal + 0.02·Spatial
Current native decoder-aware stage:
L = mean(e²) + 160·mean([clip(D(ẑLTX)) − clip(D(zLTX))]²)
D is the frozen LTX Conv VAE; pixel L1, SSIM, gradient, temporal, and perceptual weights are all 0.
Promoted composition:
θleader = θfrequency leader
+ (θnative16K,3200 − θnative8K,1600)
θleader = θfrequency leader
+ (θnative16K,3200 − θnative8K,1600)
The anchor retains scale-4, native-8K, pooled-Jacobian, channel-Gram, and decoder-frequency gains. Doubling native pairs at matched exposure adds +0.01139 dB on formal exact-32. The deployed adapter remains the same 194.76M Conv3D network.
Fast-motion quality
Sharper motion, with causal controls
Fast-action blur is evaluated separately from average PSNR. Every new treatment must improve reconstruction without regressing the frozen highest-motion quartile.
Q4
Teacher-frozen motion audit
Sixteen uniformly sampled teacher frames at 192×336 define four motion quartiles. The same mask and teacher Farneback flow are reused for parent, control, and treatment—never inferred from the candidate.
3D
Selective ≤200M scaling
The prepared treatment replaces only the final three depthwise temporal convolutions with dense channel-time mixing: 194.759M → 199.842M parameters. Its diagonal initialization is bit-exact; a matched depthwise arm isolates the capacity effect.
10%
Full-grid motion loss
A later matched A/B keeps the full 768×1344 VAE decode and adds teacher-motion-weighted moving-edge, velocity, and acceleration Charbonnier terms at 192×336. Their combined parameter-gradient budget is capped at 10% of 160×pixel-MSE.
Promotion remains fail-closed. At step 400, a treatment must beat both its matched control and the resolved parent on development; this only authorizes continuation to step 800. Formal promotion then requires fresh development, disjoint audit, and formal exact-32 PSNR/SSIM, non-regression in motion PSNR/SSIM, edge sharpness, temporal error, and flow-warp residual, plus visual review of the eight quartile-4 clips for ghosting, blur, and color instability.
Measured progress
From latent pilot to 28.78 dB
Every promoted point uses decoded validation evidence. The latest points use the identical audited 768×1344 protocol.
23.32lossless packed latent
26.53equal-exposure decoder
27.38full-grid latent
27.98audited raw pixel MSE
28.12decoder exposure
28.2816K latent + task vector
28.39native 2K pairs
28.7816K native matched exposure
Evaluation contract
What the number means
Selection and reporting are deliberately separated from training proxies.
32
Frozen validation IDs
The same exact 32 videos are paired by sample ID for every formal comparison. Candidate and baseline geometries must match before deltas are computed.
192
Full temporal extent
Every clip contains 192 evaluated frames at 768×1344. PSNR uses all decoded pixels; temporal error compares first-frame differences.
SHA
Version identity, not inference
Local and official LTX-2.5 Diffusers Conv VAE weights share SHA-256 425f0dfa…55c5 and byte size 1,452,233,194. No adapter rerun is needed to relabel the identical decoder.
Reference, not an apples-to-apples baseline: the official LTX-2.3 latent upsampler without its refiner reaches 30.0742 dB / 0.89125 SSIM on this set, but it begins from an LTX latent. The adapter solves the harder cross-VAE H3→LTX translation problem.
Running next
Closing the final 0.22 dB
Blind scale is closed: a 296.66M identity-expanded control produced no decoded gain. The remaining tests target unique-data exposure and selective temporal capacity.
65,536 latent pairs
One full exposure at the same 194.76M architecture adds +0.07897 dB with 32/32 PSNR wins.
78,400 latent pairs
Standard training adds +0.08061 dB over the prior leader with a positive exact-32 confidence interval and 32/32 PSNR wins.
3.00× direction scale
The last controlled step adds +0.04270 dB over 2.00× (95% CI +0.03781 to +0.04749), with PSNR and SSIM gains on all 32 IDs.
4,096 native pairs
The rebased native-data control adds +0.01266 dB (95% CI +0.00979 to +0.01584) and wins on 31/32 IDs.
Pooled decoder Jacobian
Replacing the older sensitivity estimate adds +0.01035 dB (95% CI +0.00728 to +0.01358) and improves PSNR on 28/32 IDs.
8,192 native pairs
The rebased direction adds +0.01249 dB on exact-32 (95% CI +0.01064 to +0.01438), with PSNR gains on all 32 IDs and SSIM gains on 31.
Frozen direction scale 4
The sole selected point passes disjoint audit and formal exact-32, adding +0.02127 dB (95% CI +0.01610 to +0.02640). The scale grid is closed.
Decoder-Jacobian channel Gram
The stable 128×128 PSD metric adds +0.00486 dB on formal exact-32 (95% CI +0.00421 to +0.00556), with PSNR and SSIM gains on all 32 IDs and no inference-size change.
Decoder-frequency weighting
The fixed 78,400-step treatment passes fresh dev, disjoint audit, and formal gates. It adds +0.00133 dB (95% CI +0.00112 to +0.00153) with 31/32 PSNR wins and no inference-size change.
16,384 native pairs
The fixed step-3,200 direction passes fresh dev, disjoint audit, and formal gates. It adds +0.01139 dB on formal exact-32 (95% CI +0.00851 to +0.01412), while SSIM adds +0.000461.
Strong-parent consolidation
The first fixed block is complete. Internal validation changes by +0.0618 dB PSNR and +0.00146 SSIM, but this is diagnostic only. The low-frequency recovery monitor will resume the remaining audited blocks; only step 3,200 can enter formal gates.
One complete native exposure
The running endpoint covers 3,200/16,384 unique native pairs (19.53%). Only if consolidation passes all three gates but remains below 29 dB, the next hypothesis will preserve its optimizer and continue to exact step 16,384, then use new untouched development and audit splits before one formal decode.
Last-three dense temporal A/B
Matched native-16K control/treatment at 1e-6 train only temporal parameters in blocks 19–21. Control has 9,024 trainable parameters; treatment has 5.092M inside a 199.842M model. An older parent showed +0.03139 dB on 32/32 IDs, but no result is claimed for the current parent.
Teacher-motion auxiliary A/B
Full-grid prediction and teacher decodes remain exact. Only the loss graph is reduced to 192×336; auxiliary weights are measured from the resolved parent's parameter gradients and the treatment is materialized only below a 76 GiB peak-allocation gate.
Motion-aware exact32 gates
Standard paired PSNR/SSIM, frozen quartile-4 motion metrics, teacher-flow residuals, and eight-clip visual review are now one promotion contract. A standard-metric win alone cannot promote a future candidate.
Decoder crop fidelity
The 12×21 crop passes the resource profile, but every bounded halo from 0 through 12 fails the unchanged worst-path full-decoder fidelity gate. Crop training is closed rather than relaxing fidelity.