BennyDaBall

AI filmmaker. Creator of the Z-Engineer model series. Rendered local, posted here first.

[BennyDaBall @BennyDaBall_OG]

More testing MiniMax H3 - this time with Turbo V4 lora head to head on my 5090. same seed, same prompt, 5 recipes cut back to back in one clip.

20 steps: 149s
8 steps turbo: 62-71s depending on attention kernel. 2.4x at the fast end with sol + sage3.

Honest take: first H3 distill that didn't entirely butcher the audio. Dialogue, lipsync, sfx all survive at 8 steps.

One catch: the sol-attn runs drift the scene a little even seed-locked. Same shot, same beats, but signage and set dressing wander. The kernel changes the trajectory, not just the speed - I'm working on it.

All local. No cloud APIs.

LoRA by larryvrh https://t.co/K18bymlOES (v4_step600 ema). comfyui convert by @drbaph.

[BennyDaBall @BennyDaBall_OG]

@pikadhu @ltx_io ComfyUI!

[BennyDaBall @BennyDaBall_OG]

LTX 2.5 20 second clip in 720p on a single 5090. @ltx_io

[BennyDaBall @BennyDaBall_OG]

@sadlemonjuice @ltx_io First go around 23 minutes lol... I fixed the graph and was able to get it down to ~80 seconds on the INT8 model with Sage Attentinon 3 and chunked DiffVAE.

[BennyDaBall @BennyDaBall_OG]

With the default graph/int8 model it took ~1416s, it was thrashing offloading to system ram. I fixed it and got it down to 85 seconds on the INT8 model... This morning I have the generation time down to about ~65s using the NVFP4 ComfyUI checkpoint I whipped up yesterday + Sage Attention 3 - https://t.co/i69ksJAdwN

[BennyDaBall @BennyDaBall_OG]

LTX 2.5 in a format ComfyUI will recognize.

[BennyDaBall @BennyDaBall_OG]

Fixed the NVFP4 LTX 2.5 checkpoint so ComfyUI will recognize it.

https://t.co/BEiIpLxZdp

[LTX.io @ltx_io]

Introducing LTX-2.5, the world model the world builds on.

One of the biggest upgrades yet to the model already powering film, robotics, and real-time workflows. Higher pixel fidelity, multishot scenes that hold together across cuts, and a pretrained foundation built to be finetuned across domains.

Plus: Diffusion Fidelity Rendering, a new rendering approach built for pixel quality that holds up frame by frame, even on a cinema screen.

This isn't a tool you rent. It's a foundation you build on, open and yours.

[BennyDaBall @BennyDaBall_OG]

https://t.co/23SGZmweE1 — PiD research by @NVIDIAAI

[BennyDaBall @BennyDaBall_OG]

Latent-Tiled PiD v1.3.0 is out: added FLUX.2 Klein support. Klein renders the latents, Latent-Tiled PiD chops it up, and PiD decodes it to a seam-free 33MP in about 18 seconds on one 5090. One dropdown, 55 render-tested sizes, up to 244MP. Tile the latent, not the pixels! @ComfyUI @NVIDIAAI

[BennyDaBall @BennyDaBall_OG]

100% crop, no upscaler tricks — that's the runway reflected in his eyes. FLUX.2 Klein 9B render → Latent-Tiled PiD v1.3.0 → 7680×4352, 33MP, ~18 seconds on one 5090. Grab it in ComfyUI Manager ("Latent-Tiled PiD") or: https://t.co/23SGZmweE1 — PiD research by @NVIDIAAI

[BennyDaBall @BennyDaBall_OG]

@NVIDIAAI

[BennyDaBall @BennyDaBall_OG]

NVIDIA's PiD decoder turns latents straight into 4x pixels - but it tops out around 4K, and past that the colors quietly fall apart.

Thing is, the paper's own math says a cropped latent looks identical to a small render. PiD can't tell the difference. so instead of pushing it past its limits, we tile the raw latents and let it do the thing it's already great at - one piece at a time, then blend the seams away.

Result: seamless canvases up to 244 megapixels from a normal-sized render. no upscalers, no re-encoding, every pixel decoded from your actual latent.

v1.2 is live on the ComfyUI registry — one dropdown, 56 sizes, all render-tested on the new v1.5 checkpoints. search "Latent-Tiled PiD" in Manager and go print a wall.

Incredible decoder by the NVIDIA Spatial Intelligence Lab — @NVIDIAAIDev this thing rips

https://t.co/NbLQdgnzaE
https://t.co/LhxbQKCL7B
https://t.co/aBuefHToue
#comfyui #Nvidia

[BennyDaBall @BennyDaBall_OG]

This week @Kijaidesign dropped experimental W4A8 quantization for @MiniMax_AI H3 - 4-bit weights, int8 activations, 12.5GB file vs 19.5GB for standard INT8. Ran the matrix on my 5090: same seed per pair, 5s + 10s clips, text-to-video and reference-to-video, under Sol-Attn (@drbaph's ComfyUI port) + SageAttention. Every clip is labeled with its exact config - judge the quality yourself.

My numbers:
- Sampling speed: a wash (6.5 vs 6.8 s/it)
- The new INT8 video VAE: ~16s faster decode per clip, and it stacks with ANY model choice
- The real win is VRAM: 19.7GB vs 26.5GB peak on a 10-second render under Sol+Sage3. That ~7GB is longer-clip headroom on a consumer card
- Same seed does NOT mean same take across quants — the trajectory forks, exactly like attention-kernel swaps

All local. @ComfyUI 0.31.1 + comfy-kitchen 0.2.28.
Weights: https://t.co/grm7787Sy5

[BennyDaBall @BennyDaBall_OG]

@n11102002 Sure! https://t.co/X36kv0xf7w

[BennyDaBall @BennyDaBall_OG]

More testing MiniMax H3 - this time with Turbo V4 lora head to head on my 5090. same seed, same prompt, 5 recipes cut back to back in one clip.

20 steps: 149s
8 steps turbo: 62-71s depending on attention kernel. 2.4x at the fast end with sol + sage3.

Honest take: first H3 distill that didn't entirely butcher the audio. Dialogue, lipsync, sfx all survive at 8 steps.

One catch: the sol-attn runs drift the scene a little even seed-locked. Same shot, same beats, but signage and set dressing wander. The kernel changes the trajectory, not just the speed - I'm working on it.

All local. No cloud APIs.

LoRA by larryvrh https://t.co/K18bymlOES (v4_step600 ema). comfyui convert by @drbaph.

[BennyDaBall @BennyDaBall_OG]

@ImpactFramesX Same. I have my production workflow gated so anything under 10 seconds falls back for quality. We'll see if that changes with more tinkering.

[BennyDaBall @BennyDaBall_OG]

Just...you know, selling flat trees here.

[BennyDaBall @BennyDaBall_OG]

Tested NVIDIA's new Sol-Attn sparse attention on MiniMax H3, local on one 5090. It skips ~84% of the attention math with basically zero quality loss — so I had Dwight explain it. 10s, native audio, Sol-Attn + SageAttention 3, ~20% faster than my old stack.