Phil2Sat

/

Qwen-Image-Edit-Rapid-AIO-GGUF

like 203

As i got no time to convert you could try this one https://huggingface.co/Arunk25/Qwen-Image-Edit-Rapid-AIO-GGUF/discussions/1#69296e82770260f96c71a397

Thank you, I have used his model. Your work is excellent.

Damn, Rocm 7.1 update yesterday... whole venv f***ed up.

BTW, definitively test z-image, my first impressions are great.

I have tested the z - image, yes, it is great, I'm waiting for the base and edit model 😄
BTW, do you use an AMD Graphic Card? How is the compatibility of the graphics card? It seems to be cheaper than Nvidia with the same memory capacity.

Yes, uhm, its cheap really cheap. got it for 50€ for a 16GB HBM2 Card.

BUT

Its a vega10 or better a MI25, GCN5 from around 2016, so no matmul, no intX at all. Most models need FP32 upcast and so on.
In my case the card is fully UNsupported by as good as everything. a simple Rocm update needs at least 2 days in recompiling packages.
no flash_attention, no sage_attention.

Total VRAM 16368 MB, total RAM 31968 MB
pytorch version: 2.10.0a0+gitded9bcd
AMD arch: gfx900:xnack-
ROCm version: (7, 1)
Set vram state to: NORMAL_VRAM
Disabling smart memory management
Device: cuda:0 AMD Radeon Pro WX 9100 : native
Using async weight offloading with 2 streams
Enabled pinned memory 30369.0
Using sub quadratic optimization for attention, if you have memory or speed issues try using: --use-split-cross-attention
Python version: 3.12.3 (main, Aug  1 2025, 10:08:07) [GCC 15.1.1 20250425]
ComfyUI version: 0.3.76
ComfyUI frontend version: 1.32.10

But now here im am back in business, got at least the most things back working.

Z-Image-Flash in 1024x1024 i needs around 9s/it, its okay for me and for my little test and learn setup.

I guess with Rocm 7.X the newer cards CDNA/RDNA will perform much better.

Sorry but I have to spam some images, this is no edit, this is just text2image, no image2image, no edit, just simple z-image workflow:
z-image with a character sheet in prompt, just changing the last part, what she does. I'm absolutely blasted and the best it needs half the time with double the quality.

# Character Profile

## Core Identity
- **Name:** Rayneshia
- **Race:** Elf (German descent)
- **Gender:** Female
- **Age:** 22yo
- **Ethnicity:** European (German descent) — **distinctly non-Asian facial structure**
- **Skin Tone:** Very fair, with luminous translucency and subtle rosy blush on her cheeks and nose.
- **Body Type:** Hourglass extreme — dramatic waist-to-hip ratio, slim figure with wide skeletal hips broad at the iliac crest, minimal body fat, clear visible thigh gap.

## Head & Face
- **Face Shape:** Round, with soft full cheeks, sharply tapered chin, delicate heart-shaped silhouette, smooth forehead, narrow refined jawline.
- **Eye Color:** #3f6297 with subtle #b5e0fa undertones.
- **Eye Shape:** Extremely large and expressive, wide-set, with long natural dark eyelashes and slight catchlight for realism.
- **Hair Color:** icy platinum blonde hair with cool ash-blue highlights.
- **Hair Style:** Ultra-long, extending far beyond the hips to the lower calves, side-parted on the left, small elegant braid running along the left temple framing the face, rest flowing naturally behind the back with soft waves.
- **Hair Texture:** Silky, healthy, straight and fine, no frizz, no flyaways, slight volume at crown.
- **Facial Hair:** None.
- **Facial Scars:** None.
- **Facial Accessories:** None.
- **Makeup Style:** None.
- **Expression Default:** Shy yet alluring, soft gaze with subtle confidence, gentle poise, lips gently upturned in a quiet, serene smile.

## Body & Physics
- **Muscle Definition:** Lean and wiry; minimal body fat but fit, with adult proportions and mature grace.
- **Chest Size:** larger than average, perfectly spherical, flawlessly rounded, zero ptosis, zero sagging, perky and uplifted with full upper-pole volume, smooth seamless contour, symmetrical, unrealistically flawless shape like polished marble or digital ideal.
- **Distinctive Markings:** Visible fine pores, subtle freckles on nose bridge, tiny capillaries under skin, slight redness at nose tip, no blemishes.
- **Pubic Hair:** None.

## Attire (Top to Bottom)
- **Clothing Style:** Futuristic military elegance — clean, geometric, uniform-inspired with subtle fantasy accents.

- **Upper Body:**
  - **Base Layer:** #2d3c5b one-piece military coat-dress, medium-weight structured fabric, matte finish.
  - **Outer Layer:** None.
  - **Details:** High standing collar with precise #ddecf4 edge piping, seamless front closure, tailored waist shaping, front hem cut high at the upper thighs, Ultra-long asymmetrical coat tails extending almost to the floor, #ddecf4 geometric chest panel inset resembling a pentagonal frame, smooth long sleeves with sharp contouring, subtle shoulder structure without padding, small side slits integrated for movement.

- **Lower Body:**
  - **Type:** No separate skirt — the short front section is the integrated lower portion of the one-piece coat-dress.
  - **Fit:** Clean and structured; front opening reveals upper thighs.
  - **Material:** Same #2d3c5b structured fabric as the upper body; continuous construction with minimal visible stitching.
  - **Belt:** None.

- **Footwear:** #2d3c5b thighhigh overknee-boots with a sleek, uniform silhouette.

quality: (8K resolution, incredibly detailed, cinematic composition, depth of field, Z-Image realism engine, highly detailed skin texture, award-winning photography, Z-Image Turbo quality).
view: (full body, head, feet, toes).

Generate a photorealistic Image with accurate lighting, natural textures, and realistic details. Pay attention to proper shadows, reflections and material properties. real-person, real-life, macroscopic facial details.

A Photo of Rayneshia, on the floor of a record store, sitting cross-legged with a vinyl sleeve on her thigh, her expression one of quiet discovery as her finger pauses on a track title, her gaze introspective, high-angle front-view.

wow!!!!
it is awesome, feel free to share more tricks like that with me.
Which platform do you usually post your works on? I have to check them out, haha.

thats not work i usually share, im just always testing promt adherance, how to get different results , using official qwen3, deepseek and openai, let especially qwen3 explain images or parts of images for me.

then i try to put this into some format text2img is able to understand and recreate this.

img2text2img i would say but manually adjusted.

actual aproach is to get two different characters into a z-image prompt. works just copy a second sheet called opponent.

but a fight scene and positioning, eye contact, crossed weapons, stance is hard.

left quarter right quarter works kinda. if i get it right ill post some examples.

this is the closest i get so far, further working.

getting closer...

btw thats the person i recreated...
Changed hair color as i like it that way a little bit more