FLUX.2-Klein, Qwen Image, and Z-Image Turbo: an 18-run ComfyUI test
I wanted a small answer to a practical question: with the three closest image models already installed on one 24 GB laptop GPU, which one can actually make a quiet hand-painted mountain-cabin shot—and then preserve an illustrated two-character scene—without help from a LoRA?
The short answer is FLUX.2-Klein. Both of its runs produced usable, correctly sized images. The two Qwen runs wrote valid PNG files, but every pixel was black. Z-Image Turbo completed quickly, but neither image met the scene brief: the frames were very soft, oversaturated, and wandered away from the cabin scene.
That is not a leaderboard. It is a record of this machine, these installed files, and this exact test.
What was tested
The test ran on an RTX 5090 Laptop GPU with 24,463 MiB reported VRAM, ComfyUI, a 1280 × 1792 portrait canvas, UniPC with the simple scheduler, and seeds 42 and 1337. Every run was submitted one at a time. The model cache was freed before switching models, so the GPU was never asked to hold all three models at once.
No LoRA, ControlNet, IP-Adapter, or reference image was used. The prompt set was supplied by 豆包. The setup, output validation, and review below were performed in the current Codex session; the visual scores are an assistant review, not a human preference study.
The original plan named three FP8 releases. Those exact files were not installed, so this is explicitly an installed-model comparison, not a substitute for that intended FP8 benchmark:
| Test label | Installed file used | Steps | Guidance | Official paired VAE used |
|---|---|---|---|---|
| FLUX.2-Klein | flux-2-klein-9b-Q8_0.gguf | 12 | FluxGuidance 3.5; sampler CFG 1.0 | flux2-vae.safetensors |
| Qwen Image | qwen-image-2512-Q6_K.gguf | 12 | CFG 3.5 | qwen_image_vae.safetensors |
| Z-Image Turbo | z_image_turbo_bf16.safetensors | 7 | CFG 3.5 | ae.sft |
FLUX has a model-native guidance node. Keeping the sampler CFG at 1.0 there avoids applying guidance twice; the effective requested guidance is still 3.5. Qwen and Z-Image used CFG=3.5 in the sampler.
The shared negative prompt was:
3d, cg, blender, octane render, unreal engine, photorealistic, plastic texture, glossy skin, sharp modern digital illustration, ugly, deformed, disfigured, extra limbs, bad anatomy, oversaturated colors, neon, modern anime, ai artifacts, noise artifacts, watermark, text, signature
Prompts
The test deliberately used model-adapted prompts rather than forcing one tag list on all three architectures.
FLUX.2-Klein
Studio Ghibli animation scene, Hayao Miyazaki style, young girl standing on wooden terrace of mountain cabin at dusk, long hair blown by wind, distant layered mountain forest, thin evening mist, soft warm sunset diffused light, hand-painted cel shading, subtle watercolor texture, faint fine film grain, quiet nostalgic atmosphere, cinematic wide shot
Qwen Image
吉卜力工作室动画画面，宫崎骏手绘风格。黄昏傍晚，少女站在山间小木屋的木质露台上，长发被晚风拂动。远方是层叠连绵的山林，山谷漂浮淡淡的暮霭。复古赛璐璐手绘动画质感，带有水粉画肌理，画面有老动画细微胶片颗粒，柔和温暖的黄昏漫射光，氛围宁静怀旧，电影广角镜头构图，色彩柔和克制。
Z-Image Turbo
吉卜力动画电影画面，宫崎骏手绘风格，傍晚，少女站在山间木屋露台，长发被风吹起，远处层层山林，山间薄暮云雾，柔和温暖黄昏光，赛璐璐手绘上色，淡淡的水粉质感，微弱胶片颗粒，怀旧安静氛围，电影构图
The six outputs
FLUX.2-Klein, seed 42
This is the stronger FLUX frame. It has a readable cabin, terrace, girl, wind-blown hair, layered mountains and valley haze. The image leans more toward a contemporary anime illustration than a literal vintage film frame, but the watercolor-like paper texture and controlled dusk palette carry the intended mood.
FLUX.2-Klein, seed 1337
This second frame changes the angle and wardrobe but still satisfies the scene. The light is warm, the proportions are stable, and the mountain depth reads immediately. It is a useful confirmation that the first result was not a one-seed accident.
Qwen Image, seed 42 and 1337
Both files decode as RGB 1280 × 1792 PNGs, but each contains exactly one RGB value: (0, 0, 0). That is a completed file operation, not an image result. The Qwen VAE and text encoder were loaded according to the ComfyUI log, and older local Qwen base smoke outputs show the same all-black failure at other resolutions and settings. I therefore marked the installed Q6 setup as invalid on this host, rather than pretending a black frame is a low score.
Z-Image Turbo, seed 42
Z-Image Turbo, seed 1337
Z-Image did return non-black pixels, and it did put a long-haired figure in a green landscape. That is not enough. Both runs are visibly soft, strongly green/cyan, and include implausible bubble or lamp-like forms. The cabin terrace is not a stable subject. For this particular brief, they fail the “clean, restrained hand-painted dusk” bar.
Scores and runtime
Scores are 1–5. — means no image exists to judge. The Qwen black frames are invalid rather than “1/5 art.”
| Model / seed | D1 style | D2 scene elements | D3 structure | D4 light / depth | D5 overall | Peak VRAM | Wall time | Verdict |
|---|---|---|---|---|---|---|---|---|
| FLUX.2-Klein / 42 | 4 | 5 | 5 | 4 | 4 | 22,152 MiB | 59.923 s | Pass |
| FLUX.2-Klein / 1337 | 4 | 5 | 5 | 4 | 4 | 21,992 MiB | 30.580 s | Pass |
| Qwen Image / 42 | — | — | — | — | — | 20,843 MiB | 127.335 s | Invalid: all-black PNG |
| Qwen Image / 1337 | — | — | — | — | — | 21,017 MiB | 85.909 s | Invalid: all-black PNG |
| Z-Image Turbo / 42 | 1 | 2 | 2 | 1 | 1 | 22,713 MiB | 36.716 s | Fail |
| Z-Image Turbo / 1337 | 1 | 2 | 2 | 1 | 1 | 23,633 MiB | 26.359 s | Fail |
The first run for a model includes loading and warm-up, so wall time should not be read as a clean throughput benchmark. The useful operational observation is simpler: all successful model loads stayed under the 24,463 MiB card limit, but Z seed 1337 came closest at 23,633 MiB.
I2I: preserving an illustrated two-character scene
The I2I source is the user-supplied third attachment: a two-character illustration. I resized it without stretching from 960 × 1280 to 1280 × 1707, then added 42 px of neutral padding at the top and 43 px at the bottom to meet the fixed 1280 × 1792 canvas. The original file was not modified. The other four attachments were not submitted to any model; using them as extra identity references would turn this into a different, multi-reference test.
All I2I runs retained UniPC, the shared negative prompt, paired VAE, seeds 42 and 1337, and denoise=0.65. They were run one model at a time. The prompts were supplied by 豆包; output verification and the 1–5 review below are from the current Codex session.
I2I-A: style preservation
The intent was to preserve the original composition, poses, and framing while strengthening the hand-painted cel / watercolor / film-grain treatment.
| Model / seed | R1 composition / pose | R2 style transfer | R3 artifact control | Peak VRAM | Wall time | Verdict |
|---|---|---|---|---|---|---|
| FLUX.2-Klein / 42 | 4 | 4 | 4 | 22,039 MiB | 56.729 s | Pass with redraw |
| FLUX.2-Klein / 1337 | 4 | 4 | 4 | 21,879 MiB | 30.465 s | Pass with redraw |
| Qwen Image / 42 | — | — | — | 20,791 MiB | 124.828 s | Invalid: all-black PNG |
| Qwen Image / 1337 | — | — | — | 20,887 MiB | 85.777 s | Invalid: all-black PNG |
| Z-Image Turbo / 42 | 3 | 2 | 1 | 22,710 MiB | 35.531 s | Fail: face / clothing artifacts |
| Z-Image Turbo / 1337 | 3 | 2 | 1 | 23,658 MiB | 26.834 s | Fail: face / clothing artifacts |
FLUX keeps the broad two-person layout and gestures in both seeds, but it redraws faces, garments, and shoes. That is adequate composition preservation, not identity locking. Z keeps the rough layout yet visibly damages faces, limbs, and clothing. Qwen’s two files again decode as 1280 × 1792 RGB PNGs containing only (0, 0, 0).
I2I-B: rain-night edit
The edit prompt asked for rain, wet surfaces, and a soft night glow while keeping the original composition and poses unchanged.
| Model / seed | E1 rain-night instruction | E2 style consistency | E3 structure fidelity | Peak VRAM | Wall time | Verdict |
|---|---|---|---|---|---|---|
| FLUX.2-Klein / 42 | 2 | 4 | 4 | 22,039 MiB | 54.749 s | Partial: rain only |
| FLUX.2-Klein / 1337 | 2 | 4 | 4 | 21,879 MiB | 30.553 s | Partial: rain only |
| Qwen Image / 42 | — | — | — | 20,790 MiB | 124.097 s | Invalid: all-black PNG |
| Qwen Image / 1337 | — | — | — | 20,854 MiB | 85.028 s | Invalid: all-black PNG |
| Z-Image Turbo / 42 | 1 | 2 | 1 | 22,684 MiB | 35.091 s | Fail |
| Z-Image Turbo / 1337 | 1 | 2 | 1 | 23,670 MiB | 26.781 s | Fail |
At this fixed denoise=0.65, FLUX adds readable rain streaks while retaining the figures, but it does not turn the bright scene into a convincing night scene or make the floor substantially wet. That makes the edit a partial pass, not a success. Z neither makes a credible rain-night scene nor preserves clean structure.
I2I prompts
FLUX.2-Klein
I2I-A: Hayao Miyazaki Studio Ghibli, hand-painted cel animation, subtle watercolor texture, faint film grain, preserve original composition, preserve character pose, preserve camera framing
I2I-B: Hayao Miyazaki Studio Ghibli, hand-painted cel animation, rainy night, heavy rain, wet surface, night soft ambient glow, preserve original composition, preserve character pose, preserve camera framing
Qwen Image
I2I-A: 吉卜力宫崎骏手绘动画，赛璐璐手绘上色，微弱水粉肌理，细微老动画胶片颗粒，严格保留原图构图，保留人物姿态，保留镜头视角，不要改变物体布局
I2I-B: 吉卜力宫崎骏手绘动画画面，场景变为雨夜，大雨，物体表面湿润反光，夜晚柔和环境光；严格保留原图构图，保留人物姿态，保留镜头视角，物体布局不变
Z-Image Turbo
I2I-A: 吉卜力宫崎骏动画手绘，赛璐璐画风，淡淡的水粉质感，微弱胶片颗粒，保持原图构图不变，保持人物姿态不变，镜头视角不变
I2I-B: 吉卜力宫崎骏手绘动画，雨夜场景，大雨，物体湿润反光，夜晚柔和微光，保持原图构图，人物姿态，镜头视角不变
What I would use next
For this machine and this animation-style T2I task, I would use the installed FLUX.2-Klein Q8 workflow as the working baseline. It is the only one here that produced two complete, presentable frames without added conditioning.
I would not continue tuning the installed Qwen Q6 base blindly. The black-frame behavior repeats outside this run, and a new Qwen test needs a verified working model/runtime combination before it is worth spending more GPU time. I would also not pick the current Z-Image Turbo BF16 workflow for this brief unless its prompt path or model pairing is revisited; its failure is visual, not just a missing file.
For this source at the requested fixed settings, I would use FLUX.2-Klein for style-preserving redraws. I would not present it as a reliable rain-night editor at denoise=0.65; the result needs a stronger editing path or a separately approved parameter sweep. Qwen remains blocked by the repeatable black-frame fault, and Z needs a corrected model/runtime pairing before another I2I trial.