像素足球游戏美术实测:Qwen、Z-Image、FLUX.2 与 Anima 的 24 图对比
这次测试只回答一个制作问题:四个本地生图模型中,谁能为竖屏足球游戏提供真正有用的源参考?它不比较泛化的“好看”,而是让每个模型分别生成角色、环境和单件玩法道具,每类固定 42、1337 两个种子,共 24 张互相独立的结果。下方每张图都能直接打开原尺寸优化文件。English version
结论很明确:Qwen Image 2512 FP8 的综合表现最好,也是唯一能稳定给出可信手工像素簇的模型。FLUX.2 Klein 的环境布局与道具几何最有竞争力,但经常更像矢量或 2.5D 插画。Z-Image Turbo 很快,却不能稳定服从这套具体美术语言。Anima 理解插画、线稿与人体,但通常输出漫画,而不是像素画。
所有结果都只是源参考,不是可直接进入游戏的 sprite、背景、假转面图或运行时资产。
测试方法
所有任务均通过受管 ComfyUI 在 24,463 MiB 的 RTX 5090 Laptop GPU 上严格串行执行,SageAttention 保持关闭。没有 LoRA、ControlNet、拼图或隐藏重试。角色分辨率为 1056 × 1584,背景为 928 × 1664,道具为 1328 × 1328。
| 实际安装变体 | 本次原生工作流 | 六次均时 | 峰值显存 |
|---|---|---|---|
| Qwen Image 2512 官方 FP8 e4m3fn | AuraFlow shift 3.1,Euler/simple,50 步,CFG 4.0 | 212.044 秒 | 23,981 MiB |
| Z-Image Turbo 官方 BF16 | AuraFlow shift 3.0,Euler/simple,9 步,CFG 1.0,空负面词 | 22.422 秒 | 24,015 MiB |
| FLUX.2 Klein 9B 社区 Q8_0 GGUF | UniPC/simple,12 步,FluxGuidance 3.5,采样 CFG 1.0 | 51.811 秒 | 22,062 MiB |
| Anima Aesthetic v1.1 官方 BF16 | AuraFlow shift 3.0,ER-SDE/simple,30 步,CFG 4.0 | 36.289 秒 | 8,103 MiB |
FLUX 的文件必须准确标注为 flux-2-klein-9b-Q8_0.gguf,SHA-256 为 dfa8908d…fe5297。这是已安装的社区 Q8_0 量化,并非原计划中的官方 FP8 检查点。Qwen 使用 qwen_image_2512_fp8_e4m3fn.safetensors,Z 使用 z_image_turbo_bf16.safetensors,Anima 使用 anima-aesthetic-v1.1.safetensors。完整字节数、哈希、提示词、24 份工作流 JSON、耗时与验证收据都保存在评测归档中。
官方资料包括 Qwen Image 2512 ComfyUI 工作流、Z-Image 官方仓库、FLUX.2 Klein 模型卡、FLUX.2 ComfyUI 工作流 与 Anima 模型卡。Anima 还对照了 ComfyUI 官方模板 和一份可复现社区工作流。
角色源参考
要求是一个原创成年足球主角,完整入画,保持清晰的左前三分之四跑动踢球姿势,并且只能出现一个冰白足球。美术语言限定深青、青色、珊瑚色与硬方形像素簇。
Qwen Image 2512 FP8


两张都通过源参考验收。轮廓、人体、足球和服装母题在缩略图下仍然清晰;像素簇有明确的手工组织感,而不是简单套一层低分辨率滤镜。
Z-Image Turbo BF16


人物本身可读,但种子 42 把足球画成了畸形椭圆,种子 1337 则直接漏掉足球。大色块只是模拟低分辨率,没有形成可信的像素簇,两张都不能作为最终像素参考。
FLUX.2 Klein 9B Q8_0 GGUF


两张的人体与足球都干净,可作为构图参考。问题是边缘与明暗更像平滑矢量插画,不能把它们伪装成已完成的像素角色。
Anima Aesthetic v1.1


种子 42 出现悬浮鞋子/足球语义混乱;种子 1337 的人体很干净,但仍是网点漫画插画。后者可以辅助宽泛角色构思,却不能证明 Anima 生成了可信像素画。
足球环境源参考
环境要求是空的竖屏球员通道,视线穿过一个切角出口到夜间球场和远处单个球门;顶部与底部必须保留暗区供真实 UI 使用,同时排除人物、球、文字、记分牌和界面。
Qwen Image 2512 FP8


这组在通道纵深、远处目标、配色与阶梯边缘之间最平衡。种子 42 的矩形墙板容易被误读为记分牌,平滑光照也让两张都不是百分之百严格的像素背景。
Z-Image Turbo BF16


两张体育场都能读懂,但都丢失了要求中的球员通道机位。它们只能算体育场氛围参考,不能通过本次布局任务。
FLUX.2 Klein 9B Q8_0 GGUF


FLUX 给出了最清晰的通道结构和 UI 安全区。种子 1337 有细小彩色标记,看起来像意外标牌;两张也都偏向扁平矢量或低模式边缘,而不是手工像素簇。
Anima Aesthetic v1.1


Anima 理解通道纵深和暗构图,尤其是种子 42;但它使用漫画线稿与连续明暗。种子 1337 还丢掉了远处球门,因此只能作为局部布局草图。
单件玩法道具源参考
道具要求是一个裂开的金色球门盾:两个不对称盾片、一个小型开放球门和四组青色碎片,缩小到 64 × 64 后仍需保留强轮廓。
Qwen Image 2512 FP8


两张都只呈现一个清楚对象,盾片、球门、裂缝与碎片的分区明确。这是后续 2D 重绘最强的一组,也是在用户批准后最适合选作 3D 重建输入的一组。
Z-Image Turbo BF16


这两张是低模 3D 渲染,不是像素画。种子 1337 仍能提供连贯的几何提示,但整组只能标记为 geometry_reference_only。
FLUX.2 Klein 9B Q8_0 GGUF


两张的部件分离、轮廓和搭建线索都很优秀,适合 2.5D/3D 几何参考;但矢量式表面仍不满足原生像素画标准。
Anima Aesthetic v1.1


种子 42 是连贯但偏侧面的扁平盾牌/球门概念。种子 1337 变成包含球场、球门、立方体和大量碎片的多物体小场景,违反单件道具要求,直接拒收。
按制作用途评分
每格是两个种子、五项 1–5 分指标的平均值:角色职责可读性、任务服从、像素语言、瑕疵控制与下游制作价值。总分是每个模型 30 个指标的平均数,不是通用画质排行榜。
| 模型 | 角色 | 环境 | 道具 | 总分 |
|---|---|---|---|---|
| Qwen Image 2512 FP8 | 5.00 | 4.60 | 5.00 | 4.87 |
| FLUX.2 Klein 9B Q8_0 | 4.40 | 4.30 | 4.40 | 4.37 |
| Z-Image Turbo BF16 | 2.60 | 3.60 | 3.30 | 3.17 |
| Anima Aesthetic v1.1 | 2.80 | 3.70 | 2.70 | 3.07 |
准确提示词
四个模型保持相同语义目标,但按各自原生提示习惯和负面词策略改写。为保持可复现性,下面保留实际送入模型的英文原文,不做翻译;种子、分辨率和参数见上表,逐样本完整工作流 JSON 已归档。
角色:四个模型的原始提示词
Qwen: Create one original adult football protagonist for Kickshatter: Pixel Cup, full body and fully inside a portrait action plate, in a clear three-quarter-left running-kick pose with exactly one ice-white football in front of the planted boot. The player has warm brown skin, a compact athletic build, short dark-auburn high-top hair with one small cyan forelock, a sponsor-free midnight-teal shirt with a single coral diagonal shard motif, dark teal shorts and socks, and pale ice boots. This must read as native 2D pixel-game art at the same intentional authored scale as a 232 by 312 character plate: crisp hard square pixel clusters, stepped diagonal edges, hand-placed-looking highlight ramps, selective deep ink outlines, and a restricted palette of Ink #020A12, Tunnel #071723, cyan, mint, gold, coral, ice-white, and warm-brown skin. Use a flat deep-ink field with only a few short cyan and coral speed strokes; leave no frame, card, typography, logo, UI, crowd, stadium architecture, or extra object. The head, kicking leg, planted leg, both boots, and the one ball must remain legible at a 25 percent thumbnail crop.
Qwen negative: photorealistic, 3D render, low-poly, isometric, CGI, smooth vector, glossy material, smooth gradient, airbrush, antialiasing, blur, poster, trading card, border, text, letters, numbers, sponsor, logo, watermark, real player likeness, club, flag, duplicate ball, second person, cropped limbs, malformed anatomy
Z-Image Turbo: One original adult football protagonist for Kickshatter: Pixel Cup, full body in a clear three-quarter-left running-kick pose with exactly one ice-white football ahead of the planted boot. Warm brown skin, compact athletic build, short dark-auburn high-top hair with one small cyan forelock, sponsor-free midnight-teal shirt with one coral diagonal shard, dark teal shorts and socks, pale ice boots. Native 2D pixel-game action plate at the authored scale of a small 232 by 312 character sprite: deliberate hard square pixel clusters, stepped diagonal edges, selective deep-ink outlines and highlight ramps, restricted ink, tunnel blue, cyan, mint, gold, coral, ice-white and warm-brown palette. Flat deep-ink field with only a few cyan and coral speed strokes, no frame, text, UI, logo, crowd, scenery, second subject or second ball; head, kicking leg, planted leg and both boots readable at 25 percent thumbnail size. Negative: empty.
FLUX.2 Klein: One original adult football protagonist for Kickshatter: Pixel Cup, full body in a clear three-quarter-left running-kick pose with one ice-white football ahead of the planted boot. Subject: warm-brown skin, compact athletic build, short dark-auburn high-top hair with one small cyan forelock, sponsor-free midnight-teal shirt with one coral diagonal shard, dark teal shorts and socks, pale ice boots. Action: a readable run-up and kick, both boots and the ball clear at a 25 percent thumbnail. Style: native 2D pixel-game action plate at the authored scale of a 232 by 312 character sprite, deliberate hard square pixel clusters, stepped diagonal edges, deep-ink outlines, hand-placed highlight ramps, restricted ink, tunnel blue, cyan, mint, gold, coral, ice-white and warm-brown palette. Context: a flat deep-ink field with a few cyan and coral speed strokes, no frame, card, lettering, logo, UI, crowd, architecture or extra subject. Negative: empty.
Anima: masterpiece, best quality, safe, 1boy, original football protagonist, full body, three-quarter left running kick, one ice white football ahead of the planted boot, warm brown skin, compact athletic build, short dark auburn high top hair, cyan forelock, sponsor-free midnight teal shirt, coral diagonal shard motif, dark teal shorts, dark teal socks, pale ice boots, native 2D pixel game art, 232 by 312 action plate scale, hard square pixel clusters, stepped diagonal edges, selective deep ink outlines, hand placed highlight ramps, restricted ink tunnel blue cyan mint gold coral ice white warm brown palette, flat deep ink field, cyan and coral speed strokes, head kicking leg planted leg both boots and ball readable at 25 percent thumbnail
Anima negative: worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, photorealistic, 3D render, low-poly, isometric, smooth vector, glossy material, smooth gradient, airbrush, antialiasing, poster, trading card, border, text, letters, numbers, sponsor, logo, watermark, real player likeness, club, flag, duplicate ball, second person, cropped limbs, malformed anatomy
环境:四个模型的原始提示词
Qwen: Create one empty portrait menu-background plate for Kickshatter: Pixel Cup. Look straight out from a deep Ink #020A12 player tunnel through one angular cut-corner opening to a quiet night football pitch and one small ice-white distant goal. The tunnel has dark Tunnel #071723 walls, deliberately stepped edges, a few tiny gold practical lights, cyan floodlight rims, deep blue stadium tiers, sparse square crowd-light clusters, and only one narrow pitch-green stripe at the far end. This must be a native 2D pixel-game background at the same authored scale as a 420 by 912 logical-point portrait screen: deliberate hard pixel clusters, stair-step perspective, discrete dithering only where needed, no blur, no smooth gradients, no glossy 3D materials. Keep the top and bottom dark enough for native game UI, while the central opening and goal stay readable. No players, ball, obstacle, menu, labels, text, logos, scoreboard, flags, camera, equipment, watermark, border or UI.
Qwen negative: photorealistic, 3D render, low-poly, isometric, CGI, smooth vector, smooth gradient, soft blur, bloom, lens flare, poster, screenshot UI, typography, letters, numbers, logo, watermark, player, person, football, obstacle, crowd faces, flag, camera, tripod, scoreboard
Z-Image Turbo: One empty portrait menu-background plate for Kickshatter: Pixel Cup. Centered view from inside a deep-ink player tunnel through one angular cut-corner opening to a quiet night football pitch and one small ice-white distant goal. Dark tunnel-blue walls, stepped pixel edges, a few tiny gold practical lights, cyan floodlight rims, deep-blue stadium tiers, sparse square crowd-light clusters and one narrow pitch-green stripe at the far end. Native 2D pixel-game background at a 420 by 912 screen scale: hard pixel clusters, stair-step perspective, restrained discrete dithering, a dark quiet top and bottom for real UI, a readable opening and goal. No player, football, obstacle, menu, text, logo, scoreboard, flag, camera, equipment, border or UI. Negative: empty.
FLUX.2 Klein: One empty portrait menu-background plate for Kickshatter: Pixel Cup. Subject: a centered view from within a deep Ink #020A12 player tunnel, through one angular cut-corner opening, toward a quiet night football pitch and one small ice-white distant goal. Style: native 2D pixel-game background at the authored scale of a 420 by 912 logical-point screen, hard pixel clusters, stepped perspective, minimal discrete dithering, dark tunnel-blue walls, cyan floodlight rims, a few gold practical lights, deep-blue stadium tiers, sparse square crowd-light clusters, one narrow pitch-green stripe. Context: upper and lower frame remain dark and quiet for the real UI; no player, ball, obstacle, menu, lettering, logo, scoreboard, flag, camera, equipment, border or interface. Negative: empty.
Anima: masterpiece, best quality, safe, original empty portrait menu background for a pixel football game, centered view from inside a deep ink player tunnel through one angular cut-corner opening to a quiet night football pitch and one small ice white distant goal, dark tunnel blue walls, hard pixel clusters, stair step perspective, a few tiny gold practical lights, cyan floodlight rims, deep blue stadium tiers, sparse square crowd light clusters, one narrow pitch green stripe, dark quiet top and bottom UI safe areas, native 2D pixel game background at 420 by 912 screen scale
Anima negative: worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, photorealistic, 3D render, low-poly, isometric, CGI, smooth vector, smooth gradient, soft blur, bloom, lens flare, poster, screenshot UI, typography, letters, numbers, logo, watermark, player, person, football, obstacle, crowd faces, flag, camera, tripod, scoreboard, border, interface
道具:四个模型的原始提示词
Qwen: Create one original split gold goal-shield obstacle for the 2D pixel-football puzzle Kickshatter: Pixel Cup. Use one readable front three-quarter gameplay view. It is a thick angular gold shield wall split into two asymmetric halves by one diagonal impact line; the opening reveals one small ice-white football goal with a dark-navy square net. Four large cyan glass shard clusters are falling from the split. This is a small 2D gameplay prop, not a model sheet or product render: give it one bold hard silhouette, deliberate square pixel clusters, stair-step corners, deep Ink #020A12 outline, three flat gold highlight ramps, ice-white goal frame, cyan shards, dark net pixels and a very low pitch-green strip under a flat deep-ink background. The split, both gold halves, the open goal and cyan shards must remain immediately readable after a nearest-neighbour reduction to 64 by 64 pixels. No player, football, stadium, pedestal, studio ground, title, number, logo, text, border, HUD, soft shadow, blur, smooth gradient, isometric view, low-poly render, photoreal material, CGI or vector finish.
Qwen negative: photorealistic, 3D render, low-poly, isometric, CGI, smooth vector, smooth gradient, glossy material, product render, studio scene, airbrush, blur, antialiasing, text, letters, numbers, logo, watermark, player, football, stadium, crowd, card frame, UI, contact sheet, duplicate prop
Z-Image Turbo: One original split gold goal-shield obstacle for Kickshatter: Pixel Cup in a readable front three-quarter gameplay view. A thick angular gold shield wall splits into two asymmetric halves along one diagonal impact line and reveals one small ice-white open football goal with a dark-navy square net; four cyan glass shard clusters fall from the split. Small native 2D pixel-game prop: bold hard silhouette, deliberate square pixel clusters, stair-step corners, deep-ink outline, three flat gold highlight ramps, ice-white goal frame, cyan shards, dark net pixels, very low pitch-green strip under a flat deep-ink background. The split, gold halves, open goal and cyan shards must read at 64 by 64 pixels. No player, ball, stadium, pedestal, studio ground, text, logo, UI, smooth shading, product render or extra prop. Negative: empty.
FLUX.2 Klein: One original split gold goal-shield obstacle for Kickshatter: Pixel Cup, viewed from a readable front three-quarter gameplay angle. Subject: one thick angular gold shield wall, split into two asymmetric halves around one small ice-white open football goal with a dark-navy square net; four cyan glass shard clusters fall from one diagonal impact line. Action: both shield halves pull apart and reveal the goal. Style: a small native 2D pixel-game prop, bold hard silhouette, deliberate square pixel clusters, stepped corners, deep-ink outline, three flat gold highlight ramps, ice-white goal frame and cyan shard highlights. Context: flat deep-ink background, one very low pitch-green strip, no studio ground, no 3D material, no player, ball, stadium, text, logo, UI or extra prop; the split, goal and shards must read at 64 by 64 pixels. Negative: empty.
Anima: masterpiece, best quality, safe, original split gold goal shield obstacle for a 2D pixel football puzzle, one readable front three-quarter gameplay view, thick angular gold shield wall split into two asymmetric halves by one diagonal impact line, one small ice white open football goal with a dark navy square net, four large cyan glass shards falling from the split, small native 2D pixel game prop, bold hard silhouette, deliberate square pixel clusters, stair-step corners, deep ink outline, three flat gold highlight ramps, ice-white goal frame, dark net pixels, low pitch green strip, flat deep ink background, readable at 64 by 64 pixels
Anima negative: worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, photorealistic, 3D render, low-poly, isometric, CGI, smooth vector, smooth gradient, glossy material, product render, studio scene, airbrush, antialiasing, text, letters, numbers, logo, watermark, player, football, stadium, crowd, card frame, UI, contact sheet, duplicate prop
验证与限制
24 张源 PNG 全部通过完整解码、预期尺寸、清单 SHA-256、重复检测与原尺寸视觉检查,24 个 SHA 均唯一。公开资源为每张图提供原尺寸、800 像素和 480 像素的 WebP/AVIF,带固有尺寸与延迟加载;所有派生文件也都完成了解码和尺寸检查,原尺寸派生图还与 PNG 做了像素域质量校验。
两个固定种子足以暴露明显的复现失败,但不能建立普适排行榜。各模型使用原生表达习惯,因此提示词保持语义一致而非逐字一致。Anima 与 FLUX.2 Klein 9B 权重采用非商业许可,这些图片继续只作评测参考,任何产品用途都需要单独复核许可。最后,漂亮渲染不等于通过:Z 与 FLUX 的 3D 感道具已明确标注,绝不冒充像素画。
如果进入下一轮受控编辑,我会选择 Qwen 种子 42 的角色、Qwen 种子 1337 的环境和 Qwen 种子 42 的道具。若目标更偏几何搭建而非严格像素语言,FLUX 的道具组是更合适的备选。任何一张都不会未经清理直接进入游戏。