QwenImage 2.1 adds KV caching, measured 2.55x faster image generation

RisingSayak · x · 2026-09-21

A technical thread breaks down how QwenImage 2.1 (7B DiT) incorporates KV caching to speed up denoising:

• Core idea: separate the fixed context (text and reference images) from target-image positions that change every denoising step; the fixed prefix is cached once so later steps get much cheaper.
• Sequence construction: Qwen3-VL text tokens are interleaved with condition-image latents, with target-image latents last. During prefill, each layer saves only prefix K/V.
• Contrast with MMDiT/FLUX2: classic MMDiT joint attention is bidirectional (FLUX2 caches only reference-image tokens), but QwenImage 2.1's information flow is directional, so the entire prefix (text + reference images) can be reused. In decode, each layer computes target Q/K/V and prepends cached prefix K/V.
• Benchmarks: on a warmed A100 with two 1024×1024 references at 40 steps, DiT time dropped from 50.57s to 19.86s (2.55x), at the cost of memory rising from 32.3G to 36.0G, which the author says can be further optimized.

Original post →

More from Multimodal

• Seedance 2.5 generates a photorealistic 30s Korean winter travel ad, full prompt included — SimplyAnnisa · 2026-09-21
• Open-source 3DGS tool LichtFeld goes end-to-end: real-time SLAM reconstruction at 30 FPS — janusch_patas · 2026-09-21
• Punch Monkey AI film hits 50M+ views as creator shares full generation workflow — HashemGhaili · 2026-09-21
• Vaporous glitch art prompt template for AI image generation — LudovicCreator · 2026-09-21
• Fully AI-generated music video made with Seedance 2.5 shows pricey MV production fading — aitrendz_xyz · 2026-09-21
• Qwen-Image 2.1 LoRA training plagued by anatomy failures, community hunts for fixes — ResidentFrame4195 · 2026-09-21

Multimodal channel →