all / g / s / fur / art / ai / 3dcg / f / rp / req / p2p / dis / lit / kaki chat
[Settings] [Home] [Catalog] [Search] [Thread list] [Stats] [Reports] [Watcher] [PMs] [Admin]
[Return] [Bottom]

Posting mode: Reply
BBCode
(for deletion)
  • Allowed file types are: gif, jpg, jpeg, png, webp, webm, mp4
  • Maximum file size allowed is 25000 KB.
  • Images greater than 255 * 255 pixels will be thumbnailed.
  • 179 unique users in the last 30 minutes (including lurkers)
    First
    [1]
    		Last
    File: 1x2apx17FLsjl5FAfdYaJvWT_comparison.png
    (7.15 MB, 2924x1667) ▶
    7501719
    No.29717 [Edit] [Report] AS
    ▶
    So when Qwen 2.1 came out recently it got a lot of general buzz for being the best-in-class local image edit model. I ran some of my guro collection through it with a "make it photorealistic" pass to see how it did and... it was alright, but not amazing. It often struggled a lot around any aspect of nudity and isn't amazing at gore.

    Well, I saw this reddit post about a new local model called "HunyuanImage 3.0" with image edit capabilities and it looked impressive so I gave it a shot and yeah... I heavily prefer.

    https://www.reddit.com/r/StableDiffusion/comments/1wzclz5/hunyuanimage_30_80b_running_natively_in_comfyui/

    I'll upload a few comparison images so you can see just how much better it is. The qwen images were generated with the standard qwen 2.1 workflow, with steps cranked to 40 and cfg to 2.0 like everyone recommends. The hunyuan images are the workflow from the reddit post with the lightest version of the model (4-bit instruct-distil).

    The prompts are slightly different, I couldn't be arsed to go redo the qwen ones with exactly the same prompt. The core of the prompts are just "Make the image a real photo", but sometimes I added a little extra description to help the models in areas where they were struggling. But I promise each image is the best representation I pulled out of around 5-10 images for qwen, and 1-4 images for hunyuan (most of them are the first shot, unless there's more to the prompt than just the base "make it a real photo")
    >>
    No.29718 [Edit] [Report]
    ▶
    File: 2efargsrgs_comparison.png
    (4.62 MB, 2624x1415) ▶
    4843327
    >>
    No.29719 [Edit] [Report]
    ▶
    File: 2jtzpvf0buqc1_comparison.png
    (3.7 MB, 2996x1270) ▶
    3881057
    >>
    No.29721 [Edit] [Report]
    ▶
    File: 64606492_p1_comparison.png
    (3.37 MB, 2660x1334) ▶
    3532181
    Hunyuan missed the center spike on this one but I'm sure that could be ironed out with further gens or a clarification added to the prompt.
    >>
    No.29722 [Edit] [Report]
    ▶
    File: 64606492_p16_comparison.png
    (3.77 MB, 2660x1408) ▶
    3954174
    >>
    No.29723 [Edit] [Report]
    ▶
    File: 65540650_p9_comparison.png
    (3.34 MB, 3572x1340) ▶
    3498591
    >>
    No.29724 [Edit] [Report]
    ▶
    File: 66247890_p3_comparison.png
    (3.69 MB, 2660x1334) ▶
    3864005
    >>
    No.29726 [Edit] [Report]
    ▶
    File: 66328578_p0_comparison.png
    (4.2 MB, 3140x1231) ▶
    4399701
    probably one of the clearest improvements. Qwen really couldn't handle this one (further prompt iterations to try and give it the prolapse did not work out well). idek what it's trying to do with the nudity. Hunyuan nailed it first prompt (i forgot to tell it to remove the censor bar, oh well)
    >>
    No.29727 [Edit] [Report]
    ▶
    File: 66480946_p0_comparison.png
    (5.37 MB, 3620x1173) ▶
    5633316
    >>
    No.29729 [Edit] [Report]
    ▶
    I've been relying on Qwen, but yes, definitely have to give Hunyuan a try.
    >>
    No.29730 [Edit] [Report]
    ▶
    File: 66633325_p0_comparison.png
    (3.29 MB, 2564x1411) ▶
    3450346
    >>
    No.29732 [Edit] [Report]
    ▶
    File: 69378180_p0_comparison.png
    (6.66 MB, 3476x1753) ▶
    6988410
    >>
    No.29733 [Edit] [Report]
    ▶
    File: 70208638_p1_comparison.png
    (3.7 MB, 2276x1554) ▶
    3877309
    ok, i have to be honest - i do actually kinda prefer the Qwen generation on this one. the Hunyuan one is more faithful to the original but qwen looks more "real".
    >>
    No.29734 [Edit] [Report]
    ▶
    File: 70658874_p3_comparison.png
    (3.13 MB, 2708x1404) ▶
    3282873
    >>
    No.29735 [Edit] [Report]
    ▶
    File: 1543909892035847809395798881048_compari(…).png
    (3.46 MB, 4100x1002) ▶
    3633315
    >>
    No.29736 [Edit] [Report]
    ▶
    File: 15871967896385489957667232874164_compar(…).png
    (3.58 MB, 2612x1403) ▶
    3752743
    >>
    No.29737 [Edit] [Report]
    ▶
    File: ComfyUI_01510__comparison.png
    (8.82 MB, 4340x1692) ▶
    9252602
    both models do pretty decently building off of generations from anisnuff, i've noticed (my favorite non-photoreal t2img model... the creator has a thread on this forums somewhere)
    >>
    No.29738 [Edit] [Report]
    ▶
    ok, posting spree done for now. i'm curious to see what kind of results other ppl get (and if anyone has the memory to run the bigger versions of the model... curious if there's any further quality improvement)

    i also see it has t2img, kinda want to try that but i doubt anything can overthrow krea2 with loras right now.

    btw, side rant, it pisses me off how both models have a severe case of "make-everyone-look-at-the-camera-unless-explicitly-told-otherwise"itis. qwen is at least a little better at respecting "eyes slightly rolled back" i've noticed.
    >>
    No.29784 [Edit] [Report]
    ▶
    File: imagen_2026-10-09_135213664.png
    [Animated PNG](372 KB, 1281x701) ▶
    381871
    >>29738
    could you try this one?
    >>
    No.29811 [Edit] [Report]
    ▶
    File: z_Hunyuan_1791564739099.png_11254357242(…).png
    [Animated PNG](1.32 MB, 1280x688) ▶
    1381701
    >>29784
    this was the best one from a few attempts
    >>
    No.29871 [Edit] [Report]
    ▶
    >>29784
    Dude!! Where did this come from!?
    >>
    No.29874 [Edit] [Report]
    ▶
    File: image.png
    (5.03 MB, 1747x1491) ▶
    5274922
    >>29811
    Amazing stuff.

    And although really not much similar, I can't help but think of this.
    >>
    No.29875 [Edit] [Report]
    ▶
    >>29811
    Amazing!!!!
    how long did it take? i would like to convert to this the whole fallen flowers series
    https://nhentai.net/g/656832/1/
    >>
    No.29907 [Edit] [Report]
    ▶
    File: d4c12cf43c3d826820cca7aa9a23547e.png
    [Animated PNG](469 KB, 800x600) ▶
    480307
    If you would be so kind.
    >>
    No.30022 [Edit] [Report]
    ▶
    File: z_Hunyuan_1791685032553.png_57899916493(…).png
    [Animated PNG](1.41 MB, 1184x880) ▶
    1480097
    >>29907
    lol, inflation's not really my thing but it's quick so here ya go

    >>29875
    dang, how i not heard of this series before? i might give converting it a shot at some point but it would take a while due to a couple factors...
    1. black and white means some additional clarification is usually needed in the prompt so it doesn't mistake details
    2. character consistency is a big problem for multiple panels, let alone multiple pages
    both kinda fixable with prompting but it means each page needs to be handled individually rather than just throwing all 400 some pages in a batch job
    >>
    No.30044 [Edit] [Report]
    ▶
    File: IMG_2345.webp
    [Animated WebP](204 KB, 1280x1808) ▶
    209000
    >>30022
    What are your setup specs? What was the prompt of the conversión you used in my example?
    I would love to do it, but i am abroad a few more weeks.
    For example, what would the prompt be for this one?
    >>
    No.30069 [Edit] [Report]
    ▶
    >>30044
    >>30022
    I have a 2x xeon 2397V4 with 48gb Ram, and a rtx 3080 with 12 gb ram.
    Could i use it to some degree?
    >>
    No.30195 [Edit] [Report]
    ▶
    File: z_Hunyuan_1791863897825.webp_9728961613(…).png
    [Animated PNG](1.62 MB, 864x1216) ▶
    1694500
    >>30044
    >>30069
    here are the prompts for all three conversions, in order:

    bisected:
    "Make the image a real colorized photo. The image depicts a woman who has been cut in half. The lower half of the woman is being fucked in the ass by a man and there is cum splashing on her ass, and her guts are spilling out of the waist. The upper half of the woman also has guts spilling out of her waist, and is being fucked in the guts by another man. She is holding her uterus that is spilling out of her lower half. There are two small bells hanging from her nipples."

    inflated:
    "make the image a real photo. the photo shows a young girl who has been inflated. her skin should be smooth and her mouth should be taped over and she should have tears running down her face."

    this one:
    "make the image a real photo. the image shows a young woman who has been split in half vertically with her guts hanging out and bells hanging on her nipples."

    (for all three, you should be able to download them and drag them directly into comfyui to get the exact workflow)

    i pretty much always just start with "make the image a real photo" to see what it gives me and then just add details where it messed up until it gets them right. most common addons are probably:
    -specifying blood/guts when converting a black and white photo
    -specifying youthfulness, sometimes it just decides to make people grandmas
    -specifying smooth skin, sometimes it fucks up the skin texture pretty bad (one of its more prominent flaws)
    -specifying eyes looking away from the camera/mouth being open... like i mentioned both this model and qwen like changing these for no reason

    i have a pretty powerful setup... 5090 with 64gb ram. cpu doesn't matter so much as long as its not awful. you only having 12GB of VRAM is a bit sketchy - this is a massive model, around 45GB. i think you should be able to handle it since it can spill into RAM (even the 5090 can't hold the whole model at 32GB), but it's cutting it close. still, i think you can at least try and see where it gets you.

    worst case scenario, Qwen 2.1 is a lot smaller and should fit on your setup just fine. I know i've been kinda shitting on Qwen this thread, but its definitely better than nothing
    >>
    No.30209 [Edit] [Report]
    ▶
    >>30195
    Amazing!!!!!!
    >>
    No.30260 [Edit] [Report]
    ▶
    >>30022
    >lol, inflation's not really my thing but it's quick so here ya go
    Do you have any idea how common this is and yet how often it's the absolute best? You're the best!
    >>
    No.30323 [Edit] [Report]
    ▶
    Can you make this
    https://boards.guro.st/ai/src/1792144587659.jpg
    >>
    No.31218 [Edit] [Report]
    ▶
    File: 1667784131677.jpg
    (271 KB, 800x1304) ▶
    278520
    Requesting this pretty please.
    [Top]
    Delete post: []
    First
    [1]
    		Last