Home Creators Posts Import Register Favorites Logout

Videos

  • msrv2_00001-audio.mp4

Downloads

Content

Hey everyone. If you watched the MSR V2 video, you saw the V1 vs V2 comparison and the character consistency tests. This post has the complete pipeline — the Krea 2 workflow that generates your reference character sheets, and the MSR V2 video workflow that turns those sheets into video clips.

Check out the video if you don't know what is this about. https://www.patreon.com/aifuturetech/posts/ltx-2-3-msr-v2-164795767?utm_source=youtube&utm_medium=video&utm_campaign=20260725

And here the demo video : https://youtu.be/Wh-e_1SJEg4

----------------------------------------

WHAT'S INCLUDED

----------------------------------------

Three files attached:

1. Krea 2 - t2i - Detailer With VLM Ver 202607.json — reference image generator

2. LTX-2.3_ref2vid_MSRV2_WithEnhancer.json — the video generation workflow

3. 20260725.zip — both workflows bundled together

1. Krea 2 Text-to-Image + Detailer (Reference Sheet Generator)

..............................................................

File: Krea 2 - t2i - Detailer With VLM Ver 202607.json

This is where your reference images come from. 34 nodes. Two-pass generation — Krea 2 Turbo creates the base image, then a detailer pass sharpens it up. This is what I use to generate the character sheets that feed into the MSR V2 workflow.

The pipeline:

- Krea 2 Turbo with qwen3vl_4b_fp8 text encoder and qwen_image_vae

- Pass 1: KSampler at 2560x1440, 8 steps, er_sde, simple scheduler, CFG 1, denoise 1.0

- Image scaled up 1.5x with lanczos

- Pass 2: Detailer pass, 4 steps, er_sde, simple scheduler, CFG 1, denoise 0.2

- PathchSageAttentionKJ for memory optimization

- Optional Krea2-realism-V2 LoRA (disabled by default — enable if you want photorealistic style)

- Image Comparer for A/B testing between passes

- SaveImageExtended for auto-naming outputs

The two-pass system matters. Pass 1 generates the full composition — character pose, outfit, background. Pass 2 runs on the upscaled image at low denoise (0.2) and adds texture without changing the composition. This is how you get clean, sharp reference images that the MSR V2 model can actually read well.

The optional realism LoRAs you can use with this workflow:

Krea2-realism-V1: https://civitai.red/models/2728365/krea2-realism-v1

Lenovo UltraReal: https://civitai.com/models/1662740/lenovo-ultrareal

Why this matters for MSR V2: the quality of your reference image directly affects the consistency of the generated video. Blurry, dark, or low-resolution reference images give the model less to work with. This workflow generates clean, well-lit character sheets that give MSR V2 the best chance of maintaining consistency.

2. LTX 2.3 Ref2Vid MSR V2 WithEnhancer

........................................

File: LTX-2.3_ref2vid_MSRV2_WithEnhancer.json

This is the video generation workflow — 98 nodes. It takes the reference image from the Krea 2 workflow and generates video clips with multi-subject consistency.

The pipeline:

- LTX 2.3 22B dev checkpoint

- Distilled LoRA (ltx-2.3-22b-distilled-lora-384-1.1.safetensors) at 0.5 for faster sampling

- MSR V2 LoRA (LTX-2.3-Licon-MSR-V2.safetensors) at strength 1 for character consistency

- LiconMSR node for character sheet input (1920x1280, 17 frames default)

- LTXAddVideoICLoRAGuide for IC-LoRA injection

- LTX2_NAG (scale 11, 0.25, 2.5) for guided generation

- PromptRelayEncode for text prompt routing

- Sampler: euler, 50 steps

- Audio passthrough from reference video

- Optional ComfyUI Video Anti-Aliasing Pack

The LiconMSR node is where you load your character sheet from the Krea 2 workflow. Set the resolution to 1920x1280 for landscape or 1280x1920 for portrait. The frame count default is 17 — increase to 41 or 65 for better quality. Higher frame counts give the model more reference frames to work with, which improves consistency across the generated video.

The LTX2_NAG node settings: scale 11, CFG 0.25, strength 2.5. These are the default values from the LTX team. The NAG (Normalized Attention Guidance) helps the model follow the reference more closely. Higher scale means stronger adherence to the reference image.

The full flow: Krea 2 generates your character sheet → load it into LiconMSR → MSR V2 LoRA maintains consistency across scenes → NAG guides the generation → you get video clips that look like the same character from multiple angles.

----------------------------------------

WHAT YOU NEED

----------------------------------------

For the Krea 2 workflow:

models/unet/Krea-2-Turbo.safetensors

models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors

models/vae/qwen_image_vae.safetensors

models/loras/krea2/Krea2-realism-V2.safetensors (optional, off by default)

Download Krea 2 from: https://huggingface.co/Comfy-Org/Krea-2

For the MSR V2 workflow:

models/checkpoints/ltx2/ltx-2.3-22b-dev.safetensors

models/text_encoders/gemma_3_12B_it_fp8_scaled.safetensors

models/text_encoders/ltx2/ltx-2.3_text_projection_bf16.safetensors

models/vae/LTX23_video_vae_bf16.safetensors

models/loras/ltx2/ltx-2.3-22b-distilled-lora-384-1.1.safetensors

models/loras/LTX2_3/LTX-2.3-Licon-MSR-V2.safetensors

Download MSR V2 LoRA from: https://huggingface.co/LiconStudio/LTX-2.3-Multiple-Subject-Reference

Custom nodes:

ComfyUI-LTXVideo — LTX nodes, IC-LoRA guide, NAG

Licon MSR — character sheet input

ComfyUI-KJNodes — PathchSageAttention, detailer nodes

rgthree — Power Lora Loader, Image Comparer, Any Switch

ComfyUI-VideoHelperSuite — video output

ComfyUI Video Anti-Aliasing Pack — optional

----------------------------------------

TIPS

----------------------------------------

Generate your reference image first with Krea 2, then feed it into MSR V2. Do not skip the Krea 2 step — a clean, well-lit reference image is the difference between consistent output and messy output.

The Krea 2 detailer pass at 0.2 denoise is subtle but important. It adds sharpness to the outfit details, textures, and small elements like tattoos or accessories. These details are what MSR V2 reads to maintain consistency across scenes.

For the MSR V2 workflow, start with 17 frames in LiconMSR for quick tests. Once you know the composition works, bump to 41 or 65 for the real render. The difference is noticeable — especially with complex outfits.

NAG scale controls how closely the model sticks to your reference. Default is 11. If the output drifts too far from your reference, push it to 13 or 15. If the output looks stiff, drop it to 9.

For landscape output, change the EmptyLTXVLatentVideo from 1280x1920 to 1920x1280. The LiconMSR resolution should match — 1920x1280 for landscape, 1280x1920 for portrait.

----------------------------------------

COMMON ISSUES

----------------------------------------

Reference image too dark or blurry — regenerate with Krea 2 using better lighting prompts. The MSR V2 model cannot reference what it cannot see in the input.

Character disappears in the first frames — this is normal for reference-to-video. The model sometimes takes a few frames to establish the character. Trim the first 10-15 frames in post.

Outfit changes between shots — check your reference image. If the outfit details are dark or blurry, the model fills in the gaps with whatever it thinks looks right.

Tattoos or small details not matching exactly — MSR V2 copies the style and shape, not the exact pattern. This is a limitation of reference-to-video in general. For exact details, train a character LoRA.

Memory overflow at 65 frames — drop to 41. The 22B model with LoRAs eats VRAM. 41 frames still gives good results.

Krea 2 detailer pass making image look over-sharpened — lower the denoise from 0.2 to 0.1, or skip the detailer pass entirely for softer reference images.

----------------------------------------

All three files attached. Download the model files from the links above before you try to run them.

Questions? Drop them in the comments.

Thanks for supporting the channel.

— Benji

Files

Comments

No comments found for this post.