Zironic/h3-attention-breakpoint-10s

H3 sparse-attention breakpoints: FL2VA, Ref2VA, and Turbo

This is a two-prompt visual stress test of the production 64 x 64 sparse-attention path in H3-Optimizations. It asks two different questions:

1. How low can video attention go before obvious temporal or structural artifacts appear?
2. How much attention is needed to preserve the same broad semantic trajectory as a 100% video-KV run with the same prompt and seed?

Those are not the same threshold. A sparse result can look coherent while depicting a meaningfully different action or object arrangement from the full-density control.

This is an N=2 case study, not a universal quality benchmark. The percentages below are useful operating evidence for these exact clips and settings, not guaranteed thresholds for every prompt, resolution, or workflow.

Results at a glance

Question	What happened in these clips
Standard FL2VA, visible artifacts	5% was badly broken. 10% removed most of the damage but still showed attention artifacts. At 15% and 20%, no obvious attention artifacts were found by visual inspection.
Standard Ref2VA, visible artifacts	Ref2VA needed more attention. Artifacts remained visible at 15%; at 20%, no obvious attention artifacts were found.
Semantic parity	Low-density clips could remain plausible while following a different trajectory from the 100% control. In the Turbo sweep, the visible transition occurred between 30% and 50%.
Turbo 8-step LoRA	Turbo was much more resilient. At 5%, the baker showed no obvious attention failure and the robot mainly showed light flicker rather than collapse.
Turbo fixed prefix	One 50% step followed by 10% still produced severe artifacts. Two initial 50% steps followed by 10% removed them in these cases.
Turbo smart ramp	50, 39.33, 28.67, 18, 10, 10, 10, 10% also removed the artifacts, replacing the hard prefix with a smooth decline.
Attention scheduling	At equal total requested video-KV budget, a descending early ladder avoided the transition glitch seen when attention dropped abruptly from six 50% steps to 10%.
Late attention	Raising only the final four steps back to 50% produced no meaningful visual recovery in these cases.

Test setup

Setting	Standard	Turbo
Conditioning	FL2VA and Ref2VA	FL2VA
Denoising steps	20	8
Sparse backend	Native INT8 64 x 64	Native INT8 64 x 64
Resolution	1376 x 768	1376 x 768
Duration	243 frames, 24 fps, 10.125 seconds	243 frames, 24 fps, 10.125 seconds
Sampler / scheduler	res_multistep / simple	res_multistep / simple
Audio in published primary arms	Original FL2VA and Ref2VA density ladders are silent; later schedule experiments include audio	Yes

The two fixed cases are:

• Baker: prompt_000, seed 5678. FL2VA uses both a starting and an ending keyframe. Ref2VA uses only the starting reference image.
• Service robot: prompt_002, seed 1234. It uses a starting reference image but no ending keyframe.

"10%" means that each video query is routed to a requested 10% of the available video key/value tiles. It does not mean that the result retains 10% of full-attention quality. Text and audio context are not reduced by this video-KV setting.

The 100% rows below use full video-KV connectivity through the same production INT8 attention backend. They are the correct topology controls for these sparse sweeps, but they are not full-BF16 FlashAttention-2 oracles. The exploratory single-clip BF16 FA2 arm is not used to support the breakpoint claims.

How to inspect the clips

Do not look only for frame-level sharpness. The most useful failure signals here are:

• objects deforming, disappearing, duplicating, or changing identity across frames;
• rigid background geometry breathing or shifting independently;
• the requested action changing or failing to complete;
• a coherent but different scene trajectory relative to the 100% control.

The baker is more strongly constrained because its FL2VA version has both first and last keyframes. The robot has only a starting constraint, so it has more freedom to diverge.

1. Standard FL2VA density sweep

The visible quality cliff is between 10% and 15% for these two prompts. The 5% clips fail severely, 10% retains smaller but visible attention artifacts, and 15% is the first tested density where no obvious artifacts were spotted.

Video KV	Baker	Robot
5%
10%
15%
20%
100% KV same INT8 backend

2. Standard Ref2VA density sweep

Ref2VA was less tolerant in these cases. Visible artifacts remained at 15%; 20% was the first tested density where no obvious attention artifacts were spotted. Unlike the baker FL2VA arm, Ref2VA has no ending keyframe constraint.

Video KV	Baker	Robot
5%
10%
15%
20%
100% KV same INT8 backend

3. FL2VA Turbo 8-step density sweep

Turbo was sampled at 5%, 10%, 20%, 30%, 50%, 70%, and 100%. It did not show the same catastrophic 5% collapse as the standard 20-step FL2VA path: the baker remained visually coherent, while the robot showed comparatively light flicker.

The more interesting Turbo breakpoint is semantic rather than catastrophic. The 5-30% clips follow one broad trajectory, while 50%, 70%, and 100% move onto the high-density trajectory in these samples. That places the observed transition between 30% and 50%, without implying that the lower-density trajectory is automatically lower quality.

Video KV	Baker	Robot
5%
10%
20%
30%
50%
70%
100% KV same INT8 backend

4. Equal requested compute, smoother early schedule

These two standard 20-step schedules request the same total video-KV budget: an average of 22% per step, or 12 percentage points per step above a constant 10% baseline.

• Abrupt: 50% for steps 1-6, then 10% for steps 7-20.
• Tapered: 50, 50, 40, 40, 30, 30, 30, 30, 20, 20%, then 10% for the final ten steps.

The abrupt change introduced a visible transition glitch. Spreading the same requested budget into a descending ladder removed that glitch in the comparison, showing that when attention is spent matters in addition to the total requested budget.

Baker: abrupt left, tapered right	Robot: abrupt left, tapered right

Individual clips: baker abrupt · baker tapered · robot abrupt · robot tapered

5. Dense tail attention did not visibly help

Starting from the six-step 50% prefix, restoring 50% attention for the final four steps produced no meaningful visual recovery in these cases. The useful additional attention was early; spending it only at the tail was not worthwhile here.

Schedule	Baker	Robot
No dense tail 50% x 6, then 10% x 14
Dense final four 50% x 6, 10% x 10, 50% x 4

Frame-slice comparisons: baker · robot

6. Turbo early scheduling: one step fails, two steps or a ramp succeed

The Turbo prefix sweep was measured against the full 50% run. The first-step-only schedule (50%, 10%, 10%, 10%, 10%, 10%, 10%, 10%) still has severe attention artifacts. Extending the fixed 50% prefix to two steps removes those artifacts in both tested clips.

The smart-ramp arm also removes the artifacts, using this schedule:

50%, 39.33%, 28.67%, 18%, 10%, 10%, 10%, 10%

That averages 22% requested video KV across the eight steps. It is not an equal-compute comparison with the successful two-step prefix, which averages 20%; the useful result is that gradually distributing the early attention also produces a clean Turbo output.

Failed one-step prefix versus smart ramp

The synchronized videos show the artifacted one-step prefix on the left and the clean smart ramp on the right.

Baker	Robot

Raw smart-ramp clips: baker · robot

Successful two-step fixed prefix

The existing synchronized controls show full 50% on the left and the successful two-step prefix (50%, 50%, 10%, 10%, 10%, 10%, 10%, 10%) on the right.

Baker	Robot

The three-, four-, and five-step prefix comparisons are also available under fl2va_turbo8_early_prefix_side_by_side.

Exact prompts

Baker - prompt_000, seed 5678

How the reference pictures align with the target video — Picture 1 (from Shot 1) is the starting keyframe and aligns with the 0.00-second mark of the target video; Picture 2 (from Shot 2) is the ending keyframe and aligns with the 10.12-second mark of the target video.

integrated_multimodal_description: [Shot 1] The opening frame preserves the subject placement, lighting, and composition established by <Picture 1>. Live-action, cinematic, a medium-wide shot frames a middle-aged baker opening the wooden shutters of a small street bakery before sunrise. The camera pushes in with small amplitude at slow speed as the baker with a calm voice (S1) sets a fresh loaf on the counter and says, <d>[English] Morning delivery is ready.</d> [Shot 2] At 00:06.400, the camera cuts to a close-up of the loaf being sliced while steam curls upward and the baker's hands remain steady. By the end of Shot 2, the motion settles into the subject positions, camera angle, lighting, and composition established by <Picture 2>.

overall_soundscape: Wooden shutters scrape open over a quiet street. Baking trays clink, footsteps cross the floor, and the crust crackles under the bread knife.

non_diegetic_music: A soft acoustic-guitar pattern at a moderate tempo is joined by sparse upright-bass notes and fades at the end.

Service robot - prompt_002, seed 1234

How the reference pictures align with the target video — Picture 1 (from Shot 1) is the starting keyframe and aligns with the 0.00-second mark of the target video.

integrated_multimodal_description: [Shot 1] Polished 3D CG, cinematic, the continuous wide shot begins exactly from the starting keyframe in <Picture 1> in the glass-roofed railway concourse. The camera trucks left at normal speed as the same small brass service robot walks forward carrying the same round tray with exactly three upright white cups. At 00:03.000 it gently tilts the tray while veering left around the rolling suitcase visible to its right, then levels the tray without spilling. At 00:06.500 it reaches the café counter on the left and carefully places the three cups down in one straight row by the end of the shot. The robot's brass construction, the count and appearance of the cups, the travelers, suitcase, and concourse remain consistent with <Picture 1> throughout the uninterrupted shot.

overall_soundscape: Layered station ambience continues beneath rolling luggage, footsteps, distant platform announcements, and the robot's quiet electric motors. Ceramic cups clink softly as the tray tilts, followed by three distinct taps on the café counter.

non_diegetic_music: Light marimba notes at a moderate tempo over a restrained electronic pulse, ending on three short notes as the cups touch the counter.

The Ref2VA baker prompt removes the ending-keyframe sentence and describes only the starting image. The complete FL2VA and Ref2VA prompt strings are also recorded in study.json.

Reproducibility and limitations

• Only two deliberately difficult prompts were tested. The observed thresholds should be validated on a broader prompt set before being treated as general recommendations.
• The baker FL2VA case has both starting and ending keyframes; the robot and all Ref2VA cases have only a starting constraint. Their robustness is therefore not directly interchangeable.
• These are visual diagnostics, not scalar quality scores. The 100% run is a same-seed reference trajectory, not proof of the uniquely correct motion or composition.
• The 100% controls use full video-KV connectivity through the production INT8 attention path. They should not be described as full-precision dense-attention oracles.
• Turbo 5/10/20/30/100 used the FP16 video VAE, while the later 50/70, prefix, and smart-ramp arms used the INT8 ConvRot video VAE. Large semantic differences are the focus; interpret small decode or texture differences cautiously.
• The original standard FL2VA and Ref2VA density sweeps are silent. Turbo and the later scheduling experiments include AAC audio. Video sparsity is expected to leave audio attention dense, but the silent arms cannot verify audio invariance.
• The source keyframe files and model weights are not bundled. Exact replay is therefore not turnkey.
• The MP4 files retain sanitized ComfyUI API-prompt metadata with prompts, seeds, model filenames, node inputs, and attention schedules. Compatible H3-Optimizations nodes and the referenced model files are still required.

Configuration summaries are available in study.json, manifest.json, turbo8_manifest.json, turbo8_early_prefix_sweep_manifest.json, fl2va_turbo8_smart_ramp_10pct_manifest.json, and the schedule-specific manifests.

Repository contents

• fl2va_* and ref2va_*: primary density and schedule arms.
• fl2va_*_vs_* and *_side_by_side: synchronized visual comparisons.
• diagnostics/, fl2va_constraint_maps/, and fl2va_casino_*: exploratory probe and routing controls; they are not evidence for the headline production-path conclusions.
• _superseded_silent/: obsolete silent Turbo renders retained only as provenance. Use the primary fl2va_turbo8_05pct/ files instead.
• fl2va_bf16_fa2/: a single exploratory baker render using BF16 FlashAttention-2; it is not a complete two-prompt control arm.

Browse every artifact under Files and versions.