Skip to main content

InclusionAIKandinsky LabImage model matchup

LLaDA-Image-Turbovs.Kandinsky-5.0-T2I-Lite-SFT

Which one is better?

In our own benchmark GraduallyBench Image v2, LLaDA-Image-Turbo is slightly ahead, but not certainly. Neither is in the arena. Neither has an API price.

GraduallyBench Image v2Our own benchmark0 to 100 · higher is better
31

LLaDA-Image-Turbo

Rank 79 of 108

27

Kandinsky-5.0-T2I-Lite-SFT

Rank 82 of 108

Ranks among the 21 image models with an arena rating and the 108 with a GraduallyBench Image v2 run; prices compared with 28 models with an API price.

Test images from GraduallyBench Image v2

Our own benchmarkThe images come from GraduallyBench Image v2, our own benchmark: every model gets the same four tasks, and two AI judges compare the images in pairs and blind.

Same task, the first image of each. The task with the biggest difference is preselected.

LLaDA-Image-Turbo: Infographic, first image from the fixed test run
LLaDA-Image-Turbo45 out of 100
Kandinsky-5.0-T2I-Lite-SFT: Infographic, first image from the fixed test run
Kandinsky-5.0-T2I-Lite-SFT27 out of 100

Result for “Infographic”:LLaDA-Image-Turbo ahead

Prompt given to both modelsCreate a square German infographic titled SO FUNKTIONIERT PHOTOSYNTHESE. Show exactly four numbered steps in this order: 1 LICHT, 2 WASSER, 3 CO₂, 4 ZUCKER + SAUERSTOFF. Use a clean editorial science style with one plant cross-section, simple arrows, high contrast, legible labels, and no other text.

What the rating looks at: Numbers, labels, and visual organization.

“Ahead”: the 95 % ranges of both values do not overlap. “Slightly ahead”: a higher value, but the ranges overlap.

Test protocol of the images

Each test attempt uses the same archived prompt and one request. The first output of that attempt counts. Repeats after technical test failures are disclosed for the affected images. Files are normalized to 1,024 × 1,024 pixels without cropping. Provider resolution, randomness, quality tier and prompt processing can still differ. The images are a practical visual test, not a reconstruction of the arena configuration.

Character and detail

LLaDA-Image-Turbo: manual, PyTorch 2.13.0, Diffusers 0.39.0, Transformers 4.57.6; inclusionAI/LLaDA-Image-Turbo@f4afc52d925bbac4e22a1c947111fc1f127e37e5; native 1024x1024, 1:1; one request per local attempt, first output; steps=4, guidance=1.0; seed=7961710876241130415; separate MPS initial-latent generator and global MPS scheduler RNG streams with the same integer seed; original Turbo scheduler, stochastic_sampling=true, uniform sigmas, shift=3; BF16 text encoder and diffusion components, FP32 denoising latent and CPU FP32 VAE decode; historical poster uses FP8 text encoder and remains unchanged; no CUDA equivalence claimed; MPS fallback permitted, actual fallback operations not measured, generated Nov 3, 2026, 9:54 AM. Source

Kandinsky-5.0-T2I-Lite-SFT: manual, local PyTorch 2.13.0, Diffusers 0.36.0; kandinskylab/Kandinsky-5.0-T2I-Lite-sft-Diffusers@c0158b5b20858d1fc02088db7ed4d805a296498b; native 1024x1024, 1:1; one request, first output; steps=50, guidance=3.5, seed=89463658088346984; mps for text encoder and denoising; cpu for native VAE decode; bfloat16 for text encoders; float32 for transformer and CPU VAE; native tokenizer truncation: CLIP 80 to 77 tokens; archived input unchanged, generated Nov 2, 2026, 8:12 PM. Source

Protocol gradually-image-comparison-v1. Prompt SHA 256: 985a75b38642a9f0e5e0c0e0c67f1b7f67780d42bd2168fa24e9a7a7a94b0da6

LLaDA-Image-Turbo image SHA 256: aac02152ef0efc45749b3d0a626e9f4114418394aeb6a92d492651c564f40fee

Kandinsky-5.0-T2I-Lite-SFT image SHA 256: 8d193e702ad1aa0fccdb45f5d1474d60e21d455fa58d13470edd2796af0f2aed

Infographic

LLaDA-Image-Turbo: manual, PyTorch 2.13.0, Diffusers 0.39.0, Transformers 4.57.6; inclusionAI/LLaDA-Image-Turbo@f4afc52d925bbac4e22a1c947111fc1f127e37e5; native 1024x1024, 1:1; one request per local attempt, first output; steps=4, guidance=1.0; seed=5156250474893981751; separate MPS initial-latent generator and global MPS scheduler RNG streams with the same integer seed; original Turbo scheduler, stochastic_sampling=true, uniform sigmas, shift=3; BF16 text encoder and diffusion components, FP32 denoising latent and CPU FP32 VAE decode; historical poster uses FP8 text encoder and remains unchanged; no CUDA equivalence claimed; MPS fallback permitted, actual fallback operations not measured, generated Nov 3, 2026, 9:56 AM. Source

Kandinsky-5.0-T2I-Lite-SFT: manual, local PyTorch 2.13.0, Diffusers 0.36.0; kandinskylab/Kandinsky-5.0-T2I-Lite-sft-Diffusers@c0158b5b20858d1fc02088db7ed4d805a296498b; native 1024x1024, 1:1; one request, first output; steps=50, guidance=3.5, seed=649868597087646690; mps for text encoder and denoising; cpu for native VAE decode; bfloat16 for text encoders; float32 for transformer and CPU VAE, generated Nov 2, 2026, 8:21 PM. Source

Protocol gradually-image-comparison-v1. Prompt SHA 256: ed65cb4e6a51e2336f8b149ad4cbb16e020f07d527ad94362e719d2dcf91d94f

LLaDA-Image-Turbo image SHA 256: 4e8939998310be86b1ad3dc0744a5e3ceee95a00a9113ee6732f9c06cb6b3618

Kandinsky-5.0-T2I-Lite-SFT image SHA 256: e102016732db9e8e463504a83ac0770ff8a3c4c77918143d0f9dae8498c79101

Product photography

LLaDA-Image-Turbo: manual, PyTorch 2.13.0, Diffusers 0.39.0, Transformers 4.57.6; inclusionAI/LLaDA-Image-Turbo@f4afc52d925bbac4e22a1c947111fc1f127e37e5; native 1024x1024, 1:1; one request per local attempt, first output; steps=4, guidance=1.0; seed=0; separate MPS initial-latent generator and global MPS scheduler RNG streams with the same integer seed; original Turbo scheduler, stochastic_sampling=true, uniform sigmas, shift=3; BF16 text encoder and diffusion components, FP32 denoising latent and CPU FP32 VAE decode; historical poster uses FP8 text encoder and remains unchanged; no CUDA equivalence claimed; MPS fallback permitted, actual fallback operations not measured; historical Hugging Face Space product-photo attempt used the same archived prompt and seed 0 but a different request body and backend; its AppError left the output unresolved; this is not an exact request replay or a numerical-equivalence comparison, generated Nov 3, 2026, 9:48 AM. Source

Kandinsky-5.0-T2I-Lite-SFT: manual, local PyTorch 2.13.0, Diffusers 0.36.0; kandinskylab/Kandinsky-5.0-T2I-Lite-sft-Diffusers@c0158b5b20858d1fc02088db7ed4d805a296498b; native 1024x1024, 1:1; one request, first output; steps=50, guidance=3.5, seed=4636200764941689006; mps for text encoder and denoising; cpu for native VAE decode; bfloat16 for text encoders; float32 for transformer and CPU VAE; native tokenizer truncation: CLIP 80 to 77 tokens; archived input unchanged, generated Nov 2, 2026, 8:02 PM. Source

Protocol gradually-image-comparison-v1. Prompt SHA 256: c67fd71e91304458a2feef5c91efa8677260fab0c0099a29a2c52cb579f5b2f3

LLaDA-Image-Turbo image SHA 256: e61fb35e0c12941d28d32b0e0125b934bc0857fc49216a7945406e3bf58b4316

Kandinsky-5.0-T2I-Lite-SFT image SHA 256: 73b31dc5a48059f7712c6a125ac4855f2401133101172c4bfbaa11b76c5ceee1

Typography and layout

LLaDA-Image-Turbo: manual, 1:1, native 1024x1024, one request in the successful technical repeat, first output; two total submissions after a local download failure; seed 0 (fixed provider default, randomize_seed=false); BF16 transformer, FP8 text encoder; 4 steps, CFG 1.0; HF Space commit 7b02431121944d65b8065873b48eec736a0c1022; source PNG SHA-256 d1866b4994a664728de07f1c1897776bbf04409b4a00ad7fa98666036a78816d, generated Nov 2, 2026, 3:03 PM. Source

Kandinsky-5.0-T2I-Lite-SFT: manual, local PyTorch 2.13.0, Diffusers 0.36.0; kandinskylab/Kandinsky-5.0-T2I-Lite-sft-Diffusers@c0158b5b20858d1fc02088db7ed4d805a296498b; native 1024x1024, 1:1; one request, first output; steps=50, guidance=3.5, seed=378697477899808767; mps for text encoder and denoising; cpu for native VAE decode; bfloat16 for text encoders; float32 for transformer and CPU VAE; three local execution attempts: first aborted on an MPS dtype error; second returned an unsaved latent but the runner read the wrong output field; accepted first saved image after correcting that field with unchanged prompt, seed and FP32 sampling settings; lost latent content remains unknown, generated Nov 2, 2026, 7:52 PM. Source

Protocol gradually-image-comparison-v1. Prompt SHA 256: 2e012c3d14ee5dd02203a94f394ca80a5038cdebd6032b25a8b8a221f2255419

LLaDA-Image-Turbo image SHA 256: de0c780506575a32050ce9ca3d3c4dd1a0c4cc7ccf7f7f2f6aaf17ad01295901

Kandinsky-5.0-T2I-Lite-SFT image SHA 256: a84280dd219692f7b20722a37491d6788d1ecdc594f816fa1203981642501ab2

Subscription and API

LLaDA-Image-Turbo

Subscription
n/a No plan in Germany names exactly this model.
Via API
No API price known (access: local).
Weights
Open weights.

Kandinsky-5.0-T2I-Lite-SFT

Subscription
n/a No plan in Germany names exactly this model.
Via API
No API price known (access: local).
Weights
Open weights.

Subscriptions only count plans that name exactly this model. Subscription prices include VAT, API prices are in dollars.

GraduallyBench Image v2 from Nov 7, 2026 · Subscription database from Sep 29, 2026 (45 days old)

All data side by side

The better value is bold. For ratings only when the gap is clear.

KennzahlLLaDA-Image-TurboKandinsky-5.0-T2I-Lite-SFT
Ratings
Arena ratingfrom votes by users, higher is bettern/an/a
Arena rankn/an/a
Arena votesmore votes, a safer ratingn/an/a
GraduallyBench Image v2our own benchmark, 0 to 1003127
GraduallyBench Image v2 by task
Typography and layoutLegible text, hierarchy, and composition1118
Product photographyMaterials, reflections, and fine details1728
Character and detailAnatomy, clothing, and spatial consistency4937
InfographicNumbers, labels, and visual organization4527
API price
Per imagelower is bettern/an/a
100 imagesn/an/a
Profile
ReleasedSep 4, 2026n/a
Weightsopen means: runs locallyopenopen
Accesslocallocal
Parameters6 billion6 billion
Largest outputnot consistently publishednot consistently published
Plans in Germanynaming exactly this modeln/an/a

API prices for one image at 1,024 × 1,024 pixels according to Artificial Analysis.

Sources: LLaDA-Image-Turbo model source (verified Nov 1, 2026) · Kandinsky-5.0-T2I-Lite-SFT model source (verified Nov 1, 2026)

Arena ratings by category

From blind votes on Artificial Analysis and LMArena; the sources are named under the card. “Ahead” means the lead is clear.

No shared arena score in the stored snapshot

The stored benchmark database contains no measurement for LLaDA-Image-Turbo and Kandinsky-5.0-T2I-Lite-SFT from the same arena and category. Older model versions are not used as substitutes.

More comparisons

Compare another matchup