topabaem commited on
Commit
301e170
·
verified ·
1 Parent(s): eb0bf42

Lead the card with the idol clip

Browse files
Files changed (1) hide show
  1. README.md +26 -0
README.md CHANGED
@@ -121,6 +121,32 @@ compression is visible.
121
 
122
  ![BF16 against 4-bit, five deterministic pairs](https://huggingface.co/topabaem/LTX-2.5-Text-Encoder-4bit-8GB/resolve/main/samples/comparison-sheet.png)
123
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
124
  ### The samples, as the vendor's pipeline actually renders them
125
 
126
  ![full pipeline renders, 1024x640](https://huggingface.co/topabaem/LTX-2.5-Text-Encoder-4bit-8GB/resolve/main/samples/sharp/sheet.png)
 
121
 
122
  ![BF16 against 4-bit, five deterministic pairs](https://huggingface.co/topabaem/LTX-2.5-Text-Encoder-4bit-8GB/resolve/main/samples/comparison-sheet.png)
123
 
124
+ ### 15 seconds, 1024x640, generated in the Space
125
+
126
+ <video controls width="100%" poster="https://huggingface.co/topabaem/LTX-2.5-Text-Encoder-4bit-8GB/resolve/main/samples/moe-idol-poster.png" src="https://huggingface.co/topabaem/LTX-2.5-Text-Encoder-4bit-8GB/resolve/main/samples/moe-idol-15s.mp4"></video>
127
+
128
+ ![poster](https://huggingface.co/topabaem/LTX-2.5-Text-Encoder-4bit-8GB/resolve/main/samples/moe-idol-poster.png)
129
+
130
+ 361 frames with sound, from a 300-word prompt naming ten sequenced gestures, made
131
+ entirely through [the Space](https://huggingface.co/spaces/topabaem/LTX-2.5-Text-Encoder-4bit-8GB-Demo)
132
+ on ZeroGPU in 198 s. The full prompt is in
133
+ [`samples/idol/prompt.txt`](https://huggingface.co/topabaem/LTX-2.5-Text-Encoder-4bit-8GB/resolve/main/samples/idol/prompt.txt).
134
+
135
+ Two things had to be fixed before a clip this long held together, and both were
136
+ in this project's own defaults rather than in the compression:
137
+
138
+ * **The sampling schedule did not adapt to length.** `DISTILLED_SIGMAS` is nine
139
+ fixed numbers; `LTXVScheduler` derives its shift from the latent's token count,
140
+ and a long clip carries several times more. Left fixed, the sample never
141
+ converges — furniture renders semi-transparent and saturation halves. Clips
142
+ past 15 s now switch automatically.
143
+ * **The vendor's second pass was being skipped**, on an unmeasured assumption
144
+ that a 16 GB card could not afford it. It can, at 10.03 GiB, and it is worth
145
+ 4.1x the Laplacian variance.
146
+
147
+ Details, the measurements, and the failures behind both are in
148
+ [`samples/long/schedule.md`](https://huggingface.co/topabaem/LTX-2.5-Text-Encoder-4bit-8GB/resolve/main/samples/long/schedule.md).
149
+
150
  ### The samples, as the vendor's pipeline actually renders them
151
 
152
  ![full pipeline renders, 1024x640](https://huggingface.co/topabaem/LTX-2.5-Text-Encoder-4bit-8GB/resolve/main/samples/sharp/sheet.png)