WarmBloodAban

/

Minimax-h3_Singularity

Like 685

20gigs for another model that seems to be vaguehype is not really interesting, and demos aren't anything otherwordly

Try it... don't like it... delete it.... works tons better for most cases in my opinion. and it's sharper at smaller resolutions. Have a good day

From my testing, at least the 34gb version is superior to every other model in almost everyway. There's only some minor issues such as sound being too impactful at times. But base models instead is the opposite here for me and gives a soft thud rather then a thud, while this model gives a heavy thud, when I wanted it to just be a thud. Other then some minor audio issues thus. This model is just Minimax H3 v1.1 imo.

From my testing, at least the 34gb version is superior to every other model in almost everyway. There's only some minor issues such as sound being too impactful at times. But base models instead is the opposite here for me and gives a soft thud rather then a thud, while this model gives a heavy thud, when I wanted it to just be a thud. Other then some minor audio issues thus. This model is just Minimax H3 v1.1 imo.

That might be the lora that's loaded. I found the Mystic's and other "Booster's" really "BOOST". I mean there's a reason why they reached out to WarmBlood offering funding for his training.

20gigs for another model that seems to be vaguehype is not really interesting, and demos aren't anything otherwordly

Try it... don't like it... delete it.... works tons better for most cases in my opinion. and it's sharper at smaller resolutions. Have a good day

"just download 20 gig duh"
it's not a 500mb lora, its a whole model

20gigs for another model that seems to be vaguehype is not really interesting, and demos aren't anything otherwordly

Try it... don't like it... delete it.... works tons better for most cases in my opinion. and it's sharper at smaller resolutions. Have a good day

"just download 20 gig duh"
it's not a 500mb lora, its a whole model

Buddy... if you're worried about 20 gigs.... then this AI thing may not be for you. C'YA!!

You say I'm goofy and go back to reddit lol, as you're the one in here complaining about a model when in reality it's a skill issue - a skill issue that could easily be remedied if you just READ the creators instructions and recommendations, reading: you know, top to bottom left to right? Groups of words forming sentences, sometimes those sentences give recommendations... but you're trying to crap on the best model out there. You said you have 4 h3 models already... well... thats nice... while you keep using that hybrid loader, and decide you ever want to try and come back... heres a tip... follow the recommendations of what the creator stated. Not sure if you know this, but I dont have a NAS and not everyone is using one, I just use my... get this... a hard drive. Whooo.... crazy right?

best model out there
quite a statement eh

best model out there
quite a statement eh

For minimax? Find me one better and I'll retract my statement. Until then... dont shit on a model you haven't even downloaded because you have a 126 gb hard drive.

minimax? yeah it's peak, singularity? i don't know, and it doesn't really call me from the previews, sounds like something fixable with good prompting

minimax? yeah it's peak, singularity? i don't know, and it doesn't really call me from the previews, sounds like something fixable with good prompting

There's a reason why he was approached and offered funding.

people with money can commit mistakes

people with money can commit mistakes

well maybe you can find someone to make a mistake and get you a 500 gb hard drive 😄

Here's a same-seed, same-prompt A/B I ran today, in case it helps anyone deciding whether to download.

Setup: 1× DGX Spark (GB10, ~121 GB unified memory), native ComfyUI H3 nodes. Both runs share the text encoder (Qwen3-VL-32B int8_convrot), the official VAEs, res_multistep + simple and BasicGuider (no CFG).

• A: official minimax_h3_fl2va_bf16 + fl2v_turbo_8step_v1.0 (via MiniMaxH3ImageToVideo)
• B: Singularity_ref2va_v1.3_int8 (34 GB) + the recommended ref2v_turbo_4step_v0.1 (via MiniMaxH3ReferenceToVideo)

Caveat: A is FL2VA and B is Ref2VA, so the same seed doesn't give the same composition. This compares against the official FL2VA, not the official Ref2VA. The 20-step run without a LoRA is the fairest pair.

What I saw:

• ✅ Faces/skin: clearly nicer on Singularity in the turbo run. The skin looks natural rather than oily and the lighting is more cinematic. At 20 steps the two were about even.
• ✅ Speed: turbo 4-step is about 35–40% faster.
• ✅ Dialogue: Portuguese came out word-perfect on both (checked with Whisper large-v3), with lip-sync intact.
• ✅ Ref2V: I tested a cartoon character reference image. Identity held in close-ups, and it followed a 4-beat action chain exactly (drawer → sleds down the fridge on a spoon → crashes into popcorn → popcorn on its head).
• ❌ Wide photoreal scene: less detail on small subjects (the gull) and visibly washed-out / greyish colour. This matches the desaturation report in #18, and the saturation numbers above back it up.
• ❌ Fast cartoon action: more motion blur than the official model, not less.
• ➖ Not reproduced: no hallucinated subtitles (0/5; my prompts end with "No text, subtitles, logos or watermarks.") and no duplicated characters.

My take: for my use it's not a drop-in replacement for the official bf16. I'd use it for realistic human/face shots, for character consistency via Ref2V, and for fast drafts. I'd stay on the official model for wide scenes and stylized animation. A small saturation boost in post may cover most of the colour issue.

Thanks to the author for sharing it.