<div data-youtube-video="https://youtu.be/P2bt83Op9bY"><iframe src="https://www.youtube.com/embed/P2bt83Op9bY" width="853" height="480" allowfullscreen></iframe></div> 
 
FLUX 3 Image Klein multi-reference workflow. Feed one to ten reference images and tell the model which one owns the character, which one owns the outfit and which one owns the scene. 
 
Main features: 
 
- Built on the distilled FLUX 3 Image Klein checkpoint (flux3_image_klein_sd_int8_convrot + qwen3vl_4b text encoder + flux2 VAE) 
- ReferenceMultiLatent packs up to 10 reference images into a single conditioning tensor; reference them in the caption as ref_image_0 through ref_image_9 
- 16 steps, no classifier-free guidance, euler/simple, denoise 1.0 
- ConditioningZeroOut lets you disable a reference slot without deleting its image, which makes A/B of reference sets quick 
- ImageCompare renders candidates side by side so you can judge drift across reference counts 
 
Suggested workflow: 
 
1. Load 1 to 10 reference images into the ReferenceMultiLatent node. Start with 1-2: accuracy degrades noticeably past 4 references. 
2. Assign jobs in the caption explicitly, e.g. "ref_image_0 is the character, ref_image_1 is the outfit, ref_image_2 is the background". 
3. Run, compare with ImageCompare, and increase the reference count only when a slot is genuinely needed. 
 
Dependencies: 
 
- flux3_image_klein_sd_int8_convrot.safetensors (distilled UNET) 
- qwen3vl_4b_int8_convrot.safetensors (text encoder / CLIP) 
- flux2-vae.safetensors (VAE) 
- ReferenceMultiLatent, ImageCompare (custom nodes) 
 
RunningHub workflow page: https://www.runninghub.ai/zh-cn/user-center/1846601088051884033/webapp?inviteCode=rh-v1111 
 
Bilibili tutorial video: https://www.bilibili.com/video/BV1btJb6LEGS/ 
 
RunningHub registration with invite code rh-v1111: https://www.runninghub.ai/?inviteCode=rh-v1111

FLUX 3 Klein Multi-Reference 1-10

<div data-youtube-video="https://youtu.be/P2bt83Op9bY"></div> 
 
FLUX 3 Image Klein multi-reference workflow. Feed one to ten reference images and tell the model which one owns the character, which one owns the outfit and which one owns the scene. 
 
Main features: 
 
- Built on the distilled FLUX 3 Image Klein checkpoint (flux3_image_klein_sd_int8_convrot + qwen3vl_4b text encoder + flux2 VAE) 
- ReferenceMultiLatent packs up to 10 reference images into a single conditioning tensor; reference them in the caption as ref_image_0 through ref_image_9 
- 16 steps, no classifier-free guidance, euler/simple, denoise 1.0 
- ConditioningZeroOut lets you disable a reference slot without deleting its image, which makes A/B of reference sets quick 
- ImageCompare renders candidates side by side so you can judge drift across reference counts 
 
Suggested workflow: 
 
1. Load 1 to 10 reference images into the ReferenceMultiLatent node. Start with 1-2: accuracy degrades noticeably past 4 references. 
2. Assign jobs in the caption explicitly, e.g. "ref_image_0 is the character, ref_image_1 is the outfit, ref_image_2 is the background". 
3. Run, compare with ImageCompare, and increase the reference count only when a slot is genuinely needed. 
 
Dependencies: 
 
- flux3_image_klein_sd_int8_convrot.safetensors (distilled UNET) 
- qwen3vl_4b_int8_convrot.safetensors (text encoder / CLIP) 
- flux2-vae.safetensors (VAE) 
- ReferenceMultiLatent, ImageCompare (custom nodes) 
 
RunningHub workflow page: https://www.runninghub.ai/zh-cn/user-center/1846601088051884033/webapp?inviteCode=rh-v1111 
 
Bilibili tutorial video: https://www.bilibili.com/video/BV1btJb6LEGS/ 
 
RunningHub registration with invite code rh-v1111: https://www.runninghub.ai/?inviteCode=rh-v1111

Files with SHA256: c93643d206658e717e0e40595412106a24140dfaa335fac60928cba20bc04472