Interested in AI video generation and open source https://nitter.cf/t.co/rItpNvosal
Joined July 2025
- Tweets137
- Following57
- Followers96
- Likes38
tazmannner retweeted
✧ 大学1年生の自主制作アニメーション｜『ECLIPSE』 ✧
四人の勇者が、女神を巡る聖なる巡礼の旅へと踏み出す。
@AiZi0102
@KTNKTN9211
@xH7NeggIQmJfkFy
@OuO_Chen25
@nicky_ni7
Added the next part and subtitles, stiched together. @pika_labs #h3 #Pika Flex + my upscale piple + my other anime nodes with refmix model.
Added the next part and subtitles, stiched together. @pika_labs #h3 #Pika Flex + my upscale piple + my other anime nodes with refmix model.
testing @pika_labs #pika H3 Flex with speed dial, 0.7 speed + my upscale 0.5m to 2mp pipeline (derope fast, animesr, face detailer, and anime cadance) + my edge runners style lora. All 1 click t2v wf. Had to create a low vram node to get it to not OOM.
github.com/matlowai/ComfyUI-…
There is something about this new Pika h3 lora and anime, the lines seem a lot cleaner, wideshots with faces less mushy and mistakes often made by other turbo loras no present.
Honored to win 5th place in Civitai's lora training contest, I'll keep working on improving on style loras 🩷
civitai.red/models/2949009/r…
Tested new Sol-H3-VAE and its x2 faster on the decode but with noticeable gridlines with fast motion. So its not worth the 9% speed up on a 2mp upscale pipeline. You could stick it in the first pass with no negatives but that just gives you a 1% speed up.
github.com/Comfy-Org/ComfyUI…
Updated my #H3 Hires Relay node to include pika workflows
github.com/tazmannner/ComfyU…
More @pika_labs #Pika H3 Flex tests. Default 1mp comparison to DMAD. Pika is the winner, dmad euler beta57 is not bad but it has a lot shift from the base model 20 steps version.
Tested @pika_labs Pika H3 Flex. And this is now the new king of turbo loras hands down. And with my H3 Hires-Relay it drops a full 2mp upscale to ~180s total.
With 1 pass Pika, you lose the original take but the take that Pika creates itself is not bad at all. And motion seems to handle well which is a first for turbo.
You can compare audio, you can see there is a difference (like with visuals different) but its not bad change, just different.
Haven't done much training lately, I guess I just have a lack of motivation. I need to sit down and plan out what I want to do next
Ran some tests on Open-H3-Regenerate-2K (2k refiner) and for anime I'm not impressed. Too many changes. For IRL content it may be a different story. (A which is my method seems best and closer to baseline)
