Krea 2 Camera Angles: Default and Two-Phase (euler) Workflow

One workflow, two ways to run the Camera Angles LoRA on Krea 2: the default (LoRA on every step) and a two-phase option (LoRA on the first steps only). It opens with the default active and the two-phase group bypassed, so one Queue renders one picture. Switch groups when the default does not hold the angle. Core ComfyUI nodes only, no custom nodes.

Required models

• Diffusion model: RedCraft 3.0 Krea 2, the INT8 convrot build, redcraft23_30Krea2_int8_convrot.safetensors, in models/diffusion_models. Plain Krea 2 and Krea 2 Turbo load too; RedCraft is what we tested.
• Text encoder: the Qwen3-VL 4B text encoder from the Krea 2 downloads (see the Krea 2 guide on comfyui-wiki.com); we run a community int8 ConvRot build of it (qwen3vl_4b_int8_convrot), and the bf16 or fp8 file works the same. It goes in models/text_encoders, loaded with CLIPLoader type krea2.
• VAE: the Qwen-Image VAE, qwen_image_vae.safetensors, in models/vae, from Comfy-Org on Hugging Face.
• LoRA: Camera Angles for Krea 2, version v2, file CameraAngles_Krea2_v2.safetensors, in models/loras. The workflow does not work without it.

The six trigger phrases

Start the prompt with one of them, in the trained short form: <phrase>, a photograph of <one sentence> or <phrase>, an illustration of <one sentence>. Keep the prompt short; long prompts fight the LoRA.

• low angle shot
• worm's eye view
• low angle shot from behind
• high angle shot from behind
• eye level shot
• high angle shot

Settings

• Steps 8. It is a distilled model, more steps do not help. CFG 1.0, scheduler simple, 1 megapixel sizes (896x1152, 1152x896, 1024x1024), tiled decode (512, overlap 64), negative a zero-out (a negative prompt does nothing at CFG 1.0).
• A. Default: LoRA at 1.0 on all 8 steps, sampler er_sde. Clean on illustration prompts and on photo prompts. Saves CamAngles_A_single.
• B. Two-phase: a first KSamplerAdvanced runs steps 0 to 2 with the LoRA at 1.3 and leftover noise on; a second KSamplerAdvanced runs steps 2 to 8 on the base model without the LoRA, with add_noise off. Same seed on both, sampler euler on both. Saves CamAngles_B_twophase.

When to use B

• Photo prompts where textures fry (skin and fabric turn crunchy) at strength 1.0.
• An angle that will not hold on A.
• Stacking with a character LoRA: put that LoRA on the model line before the sampler, so it stays on for both phases and only the camera LoRA is dropped after step 2.

B must stay on euler. We measured that er_sde across a sampler split gives colour confetti, even when the same model is loaded on both halves. The euler split is clean. On our photo test (seed 4242, LoRA 1.3 on phase 1) the euler two-phase render also gave a stronger angle than the single-phase one. Leave the sampler on euler in both nodes, keep the two seeds equal, and keep the split at step 2 on both nodes.

How to switch

• The file opens with group A active and group B bypassed (purple). To use B, select group B and press Ctrl+B to un-bypass it; bypass group A the same way if you only want the B picture. If both are active you get one picture from each in a single Queue.
• Put the same seed in the KSampler of A and in both KSamplerAdvanced nodes of B. They ship as 4242, fixed.

Credit

The two-phase idea comes from CivitAI user cutetodeath78409597, in a comment on the Camera Angles model page: in Forge they run the LoRA at 1.3 for the first 25% of steps and 0 after, and found it keeps the angle and stops photo textures from frying. Thank you for the idea.

Tested on ROCm (a 7900 XTX), about 15 seconds per picture. Core ComfyUI nodes only, so CUDA builds should run it as is.

If this worked for you, a thumbs-up here helps the next person find it.