💠💀 DaSiWa-WAN 2.2 I2V 14B | TrueVision | Safetensors 💠💀

I don't know why but, I'm getting blurry result with the main DasiwaWAN22I2V14BTruevision_boundbiteHighV10.safetensors and DasiwaWAN22I2V14BTruevision_boundbiteLowV10.safetensors. I made sure everything was in the right folder, downloaded your latest WF and updated comfyui to be sure. Works great with the GGUF but not the main model :(

@yannberg240398 This is without distillation, you need the correct settings for cfg, steps, ...
You maybe confuse this with the lightspeed version.

@darksidewalker oh thanks I will try that, i always used 4 steps and cfg 1 with the synth seduction model and didn't hae this issue. EDIT: Ah yeah I see, I was using the lightspeed model like you said. nevermind, the error was between the keyboard and the chair. XD (Me)

@yannberg240398 SynthSeduction has also this truevision and the lightspeed, you may confuse this ;)

@darksidewalker yep you're totally right, thanks a lot! :)

quick unrelated question; what do you use for your image generation?

@yannberg240398 SwarmUi with my checkpoints :)

@darksidewalker SwarmUi is the name of the workflow?

@yannberg240398 The app is swarmui

I tried your checkpoint collection, they are all very good. Will probably start generating more images. I've been mostly using my real2anime (check it out if you'd like :D) worklfow plus a custom qwen edit workflow for my work and of course your workflow for videos :P.

I have seen this myself an any wan 2.2 model, but why it is happening... I assume it is the VAE. The only thing you can do is adjusting parameters and resample till it is not happening.
For me it helps to get higher in seconds, using color-match or just use other seeds.

Have you updated comfyui to 18.1, check the changelog Releases · Comfy-Org/ComfyUI
I was having light shift issues, but more in the first frames, and I believe this fixed it., could help with your problem.

Well I tryed both but nothing changed.Still flickering and its really annoying... Thank you anyway.

@hellbergdk2 I noticed that this is likely to happen if the last frame is identical or almost to the first one. It is still happening with every WAN 2.2 model I tried. So it is most likely the VAE.

This is a fact with chaining standard WAN 2.2 videos with last frame.
Alternative is SVI 2.0

Yes, SVI 2.0 pro loras (and corresponding nodes awailable in ComfyUI manager) make miracles.

Use your initial image for color matching when stitching your video together.

Fp8+ refers to fp8 mixed precision and not just fp8 simple scaled.
All are available, just look at my collection or frontpage.

@darksidewalker I see 21 models listed in your profile page but I am not seeing a separate listing for Q8. Can you tell me how to find the Q8 version?

@ss9999 https://civitai.com/models/2456413

@darksidewalker I thought the most recent model was intended to not have the accelerators. That's only available in FP8+, then, it seems.

@ss9999 I don't get what you are onto. The recent model is available as fp8, nvfp4, gguf and without distillation as only fp8 (TrueVision)

@darksidewalker i think that the user also possibly hinted to have this undistilled version with various quantizations as well. a nvfp4 of this undistilled version would be truly a gamechanger since you are the first one to offer a nvfp4 wan 2.2 merged finetune on civitai in general iirc with the latest release. it would certainly help with svi pro workflows for quick testing purposes together with quick motion speeds right out of the box.

@darksidewalker The TrueVision model is not available in Q8. That's what has been clarified.

@JouzaDouza7135 this is a huge effort and I don't think this is the proper audience for the checkpoint, since if you use high compression like nvfp4 or gguf you are not the one who should use that checkpoint, since quality suffers. The whole matter of an high quality checkpoint destroyed by compression makes no sense to me. Also even this checkpoint is only use by like 1-2k ppl, that's not much for the effort to provide all these different quantizations. Also Civitai lacks any upload API or tool making it extremely difficult to provide all this.

@ss9999 yes, because it makes no sense. See above.

@darksidewalker I'm confused. I thought it's supposed to be higher quality because it doesn't have the accelerators. So, it seems counterproductive to not use Q8, which is supposed to have higher quality than FP8. I know you said it's FP8+ but is that equivalent or better than Q8? My understanding has been that FP8 is faster than Q8 but has worse quality (particularly in terms of prompt adherence when compared to uncompressed or lower compression like BF16 or FP16). If FP8+ mitigates the drawbacks of Q8 to the point of making Q8 pointless then, also, why offer it at all — for any of the models?

@ss9999 You are for sure confusing things here.First, here is a comparison:
https://civitai.com/articles/27174/comparison-safetensors-fp8-vs-gguf-q8-q2-vs-nvfp4
If you are not familiar why some quants exists:
https://civitai.com/articles/20293/darksidewalkers-wan-22-14b-i2v-usage-guide-definitive-edition

@darksidewalker You're the first person I've seen claim that FP8 is superior to Q8 (for aesthetic quality and prompt adherence, when compared with less compression, i.e. BF16/FP16). I have also seen comparisons from various models and Q8 has always been better than FP8, at least for similarity to BF16/FP16. If this "FP8+" is better than Q8 that's great, though, since it will be faster on Blackwell in particular than Q8. But, again, if it's better I'm not sure why you're offering Q8 for any of your models. That is, unless FP8+ needs substantially more VRAM than Q8 or requires something like Blackwell or Ada. My understanding is that plain FP8 uses roughly the same amount of VRAM as Q8. Anyway, as far as me being confused, I'll be content in that confusion because I don't think this conversation is going anywhere.

@Darksidewalker "Also gguf has the ability to offload in other ways than safetensors, what makes it good for vram poor setups." So, which is better for quality and 12 GB VRAM: Q8 or mixed FP8? The data I saw for perplexity for Q8 indicated it's very low, even for more difficult portions. Are there any similar charts anywhere that show the data for mixed FP8? I tried a Google search but didn't find anything.

@Darksidewalker I tried Q6 and Q5 when I was testing Chroma or Flux1D (reduced censorship version) or both and Q8 was faster. Multiples of 2 are faster (Q4, Q8 in particular) because of the GPU design. FP16 with both of those models was quite a bit slower than Q8. I didn't use a bunch of loras, though. I did read that using more than one lora has a speed penalty with GGUF. Three loras really slows things down. That's an issue with Wan because it needs a lot of loras to compensate for the lack of knowledge of anatomy and related matters. What I am wondering is why there are no Q8_1 quants anywhere because 1 should be better than 0.

Not gonna happen, there is already a gguf of my checkpoints and this yields no benefit.

@darksidewalker This yields a very clear benefit with 3KSamplers.

@darksidewalker  for a very dull reason. 4070 Super.

The workflow I've stuck to almost exclusively uses vanilla wan2.2 at Q6 at 720p (ish) :
https://files.catbox.moe/wdtucc.png (kill VPN if catbox gives you shit)
I only do I2V.

I started off with 2 samplers with lightning loras like everyone - 12-15 minutes for a 6s result. Then I started looking at 3Ksamplers as the motion was too limited. Thought what the heck, added 3 base high steps. Still decently fast, ended up increasing the resolution to 850p - 20 minutes tops. Movement looks much better. Finally somewhat listens to prompts. The only thing I struggled with sometimes was balancing between face consistency and accurate motion with some specific loras.

That was my go to for a while until recently - currently running a batch with 5 base steps but reverted back to 720p as going higher resulted into 2x slowdown (that I didn't have before but that's beside the point). 20-25 minutes per 8 second clip, 16FPS interpolated to 32 with GIMM-VFI. No obvious looping, and without RifleXRoPE!

Fuck yeah?

I'd say I'm loving life with this setup right now - but still getting a few terrible results here and there (genitalia related for the most part), so started looking at fine-tunes. Sure, I can go back to 2 samplers and test, but having options is always nice.

At the end of the day, GGUFs are meant to save VRAM at the expense of speed - I can see the reasoning behind not wasting time on Q4 (even Q5?) and below. But Q6/Q8 (Q6 especially) are a big deal since a lot of us opted for faster cards that unfortunately are gimped for memory at 12GB.

All I'd be hypothetically asking at this point is a fine-tune without lightning baked in at Q6 at least.

Food for thought

@darksidewalker #4 says more about civitai than anything else if anything.

I’ve tried undistilled fp16 with another workflow (my last post) as well and it works, just takesforever for obvious reasons (I only have 32GB ram so swap file gets abused).

Anyhow my main takeaway was having a quant for high noise without accelerators baked in. Whether it’s undistilled or not - tbh doesn’t matter. I concur with #5 though. Will give distilled a go.

So far I’ve settled with for smoothmix for high noise (latest does not have light2x built in) and v10 of yours for low noise. Works more than okay so far, with occasional excessive movement. Will post a few things once I get a couple batches done.

Wanted to come back about this. Thought I'd give truevision a go. For the sake of not changing a ton I left my 3KSampler intact. So apparently Ada/Blackwell makes safetensors incredibly faster than GGUF even when you have to cram 18GB into 12GB 4070Super (had to enable --lowvram).

As a result it takes the same amount of time (if not less). :shrug: You live, you learn.

@darksidewalker I set steps(25-30) and CFG(3.5) as it stated. It wouldn't be 20 mins with 4-6 steps. Mb its bcz sage_attention? I compile it myself for cuda130 for 3d models, mb i messed up. Idk

UPD. I think something with nodes or workflow. Its work fine in test workflow but make noisy mess in AiO

Not sure if you ever found solution for this but I also have been encountering same issue on steps = 30, cfg = 3.5 using the "FastFidelity C-AiO v8.2" i2v that's referenced in this model. I do have sage attention installed.

Issue seems more common on increasing resolution. Also FLF2V also seems to increase issue.

It seems to work perfectly on 360p with FLF2V.

@Darksidewalker Do I still use the SVI loras with this model when running your SVI workflow? What about lightspeed?

@Darksidewalker Thanks for the fast response. Think I edited the comment after you replied but what about LightSpeed?

@Darksidewalker Alright, thanks man. Wanted to make sure since it is labeled "LightSpeed" so wasn't entirely sure if it was baked in or not. I appreciate the work you do.

You'd think the community could think of more unique names in order to avoid confusing people... but NO, they just HAD to choose LightSpeed, LightX2V, and LTX (created by Lightricks), yet all are and mean different things.

I would say: No. I have a RTX5080 and even i use the boundbite lightspeed v10 checkpoint, it has an baked in speed lora. Truevision is just way too slow because u need like 12-20 steps and cfg higher than 2. And with the boundbite checkpoint you get really good results already with a total steps of 6-10 and an high resoloution.

@teeay25225 Thanks!

you can run full fp16 with 12gb vram 32gb ram (if u switch to linux, use interntal gpu for monitors), optimal setup for running any full model that is pubished here: 64gb ddr4 or above (96gb ddr4 or ddr5 best) , i7 or i5 (a bit slow) nvme high speed. all u need to do is configure your python env. with triton, sage-attention and use advanced diffusion loader node. i've done full fp16 and um_t5_xxl fp32 on rtx3080. your 32gb ram is it ddr5 or 4?

you might be able to run it more stable in WAN2GP

Yes you can run it. But it'll be slow.

i have the same rtx card and 12gb vram. runs fine for me. also have 128 system ram. so might be the difference