Shockwave Qwen

Shockwave Qwen is a heavily developed Qwen Image checkpoint built for high-quality, versatile image generation without requiring additional LoRAs to achieve its intended look.

Shockwave began as a long-running personal merge project and evolved through extensive experimentation with numerous Qwen Image models and derivatives. Development began before Qwen Image 2512 and later incorporated 2512-era capabilities while retaining characteristics from earlier generations. The goal was not to specialize the model around a single style, but to create a strong general-purpose Qwen Image checkpoint capable of handling a wide range of subjects, environments, clothing, lighting conditions, compositions, and photographic styles.

What Shockwave is good at

Shockwave is particularly strong with photorealism, people, clothing and outfit interpretation, complex lighting, interiors, environmental detail, multi-subject compositions, cinematic scenes, and natural photographic rendering. It can also move well beyond photography when prompted for other styles.

The model was designed to produce strong results without a LoRA. LoRAs remain fully usable when you want additional styles or capabilities, including both older Qwen Image LoRAs and newer 2512-era LoRAs.

My preferred baseline is:

8 steps • CFG 2.5 • No LoRA

Shockwave has also proven unusually flexible across CFG values and can produce strong results outside those settings, so experimentation is encouraged.

Available versions

BF16 — Full-precision master release. Recommended when memory isn't a concern.

NVFP4 — Much smaller quantized release intended for systems with more limited VRAM. Despite the substantial reduction in size, Shockwave retains remarkably strong image quality in NVFP4. Several showcase images were generated directly with this version.

FP8 — Planned as an additional middle-ground option.

Recommended components

For the results shown in the showcase, I recommend using the JIBS/Wan 2.1 2× image VAE:

Wan2.1_VAE_upscale2x_imageonly_real_v1.safetensors

An NVFP4 Qwen 2.5 VL 7B text encoder has also been successfully tested:

qwen_2.5_vl_7b_nvfp4.safetensors

Shockwave does not require an acceleration or style LoRA. A compatible 4-step LoRA can be used when generation speed is preferred, but I personally recommend running Shockwave natively at 8 steps for its intended output.

A note about the model

Shockwave wasn't created from one simple merge. It is the end result of a long experimental lineage involving well over 100 iterative merges, comparisons, back-merges, and refinements across multiple generations of Qwen Image models.

At some point I stopped counting. 😂

Rather than optimizing around benchmark numbers, I kept the versions that produced the images I wanted and continued refining from there. The result is the checkpoint you're downloading now.

Try it without a LoRA first. Give it a detailed prompt, start around 8 steps / CFG 2.5, and see what Shockwave does on its own.