dzmdd's picture
Upload folder using huggingface_hub
e488ce7 verified
Raw History Blame Contribute Delete
4.04 kB
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Do not drink the latents: H3 guide + reference evidence</title>
<style>
:root { color-scheme: dark; }
body { background:#101014; color:#e8e8ea; font:16px/1.55 system-ui, sans-serif;
max-width: 1100px; margin: 0 auto; padding: 24px 16px 80px; }
h1 { font-size: 1.5rem; margin-bottom: .2em; }
h2 { font-size: 1.15rem; margin-top: 2.2em; border-bottom: 1px solid #2a2a33; padding-bottom: .3em; }
a { color:#8ab4ff; }
p.sub { color:#9a9aa5; margin-top: 0; }
img { width: 100%; border-radius: 8px; background:#000; display:block; }
.cap { color:#9a9aa5; font-size:.92rem; margin: .5em 0 0; }
table { border-collapse: collapse; width: 100%; margin: 1em 0 .4em; font-size: .95rem; }
th, td { text-align: left; padding: .45em .8em; border-bottom: 1px solid #2a2a33; }
th { color:#9a9aa5; font-weight: 600; }
code { background:#1c1c24; padding:.1em .35em; border-radius:4px; }
</style>
</head>
<body>
<h1>Do not drink the latents</h1>
<p class="sub">Evidence sheets for a MiniMax H3 packing question: when a keyframe
guide and a Ref2VA reference are used on the same conditioning, where does the
guide actually land? Two sheets, twenty cells at one seed: four ComfyUI builds
by five reference configurations, sixteen of which render and four of which
raise. Kept here rather than on the front page because it is a note for
one upstream discussion, not part of the Motion Lab.</p>
<h2>The setup</h2>
<p>Guide rows are positioned from the text span, while reference blocks are
packed afterwards and advance the cursor the target video sits on. So every
reference slides the guide earlier relative to the clip it is meant to anchor.
The size of the slide is the reference's own cursor advance:</p>
<table>
<tr><th>reference</th><th>guide lands at</th><th>visible?</th></tr>
<tr><td>none</td><td>+0.0 units</td><td>on target frame 0</td></tr>
<tr><td>one image</td><td>&minus;1.0 unit (&minus;0.6 frames)</td><td>sub-frame, no</td></tr>
<tr><td>107-frame video plate</td><td>&minus;150 units (&minus;90 frames)</td><td>outside a 73-frame clip</td></tr>
</table>
<h2>Does the guide land?</h2>
<img src="frame0.png"
alt="Grid of generated frame 0 across builds and reference configurations">
<p class="cap">Row 1 is the input for each column. Below it, generated frame 0
measured against the guide image (mean absolute error, 0&ndash;255). The image
reference columns hold the guide everywhere, because the slide is under one
frame. The video plate of the same scene is where it breaks: 33.91, and the
beaker in the guide is simply gone. Stock master has no row because it raises
<code>RuntimeError: shape mismatch</code> on all four reference configurations.</p>
<h2>Does the reference do anything?</h2>
<img src="frame72.png"
alt="Grid of generated frame 72 across builds and reference configurations">
<p class="cap">Frame 72 of 73. Swapping an image reference for a restyled
version of the same subject moves the output by about 12/255; swapping the
video plate moves it by about 51/255, and what transfers is framing, scale and
staging rather than identity. A single seed and a single subject, so read these
as a direction, not a law.</p>
<h2>Reading the numbers</h2>
<p>Mean absolute error on 0&ndash;255 RGB, full frame. About 1.7 is seed-level
noise between two runs of the same configuration, about 5 is a cleanly anchored
frame, and a guide the model ignores lands near 70. The fix that produces the
clean rows is <a href="https://github.com/matlowai/ComfyUI/tree/fix/minimax-h3-keyframes-refs-compose">one
small branch</a>: derive the guide origin from the same reference advance the
target uses. It renders pixel-identical whether applied to master or on top of
the upstream PR.</p>
<p style="margin-top:3em"><a href="../">back to the Motion Lab</a>
&middot; <a href="https://github.com/matlowai/ComfyUI-MAINodes">repo</a>
&middot; MIT</p>
</body>
</html>