EasyAnimateV5 Sampler for Image to Video

EasyAnimateV5_I2VSampler is the workhorse of the pack: image-to-video on the current V5/V5.1 weights. Load a model, drop in a still, get a video. It's the node behind the firework demo in the official README, and the one you'll reach for when you have a frame you like and you want it to move.

The mechanism should feel familiar if you've used the legacy I2V sampler - it inherits from it - with the V5 upgrades layered on: the Flow scheduler, TeaCache speedup on by default, and a 49-frame ceiling at 8fps (about six seconds). The model is a DiT with a temporal VAE and a Qwen2-VL text encoder; the -InP weights (inpaint-style, using a CLIP image encoder) are what make still-image conditioning work.

The inputs that matter

• easyanimate_model - from the loader, with EasyAnimateV5.1-7b-zh-InP or -12b-zh-InP (or the V5 line). model_type stays Inpaint for this one.
• prompt / negative_prompt - STRING_PROMPT, from the pack's text boxes.
• start_img (optional) - your first frame. end_img (optional) - a target final frame; the model will try to land there. The end-image option is a genuinely nice EasyAnimate feature and most people never use it.
• video_length - 1–49, step 4, default 49.
• base_resolution - 512/768/960/1024, default 768. The node reads your input image's aspect ratio, picks the closest match at that resolution, and rounds to /16. You pick the quality tier; it picks the framing. The README's memory table is worth a glance here - 12B at 768×1344 needs 40GB even with offload tricks, while the 7B is comfortable on 24GB.
• scheduler - defaults to Flow, which is what V5.1 wants. The other five are legacy options.
• teacache_threshold (0.1) and enable_teacache (true) - the speed switch. Leave on; dial the threshold down if quality dips.
• seed, steps, cfg - 25 steps, CFG 7. Standard.

The output

images - an IMAGE frame batch, ready for VideoHelperSuite to write as video.

Practical notes

Same VRAM song as the rest of the 12B stack: on 24GB expect to lean on model_cpu_offload_and_qfloat8, and the 7B is the path of least resistance. And set expectations honestly - EasyAnimate's motion reads as conservative even at V5.1, so don't expect your astronaut to backflip. What you do get is strong first-frame preservation: the model sticks to your image, which makes it a much better "animate this specific thing" tool than a freewheeling T2V roll. That's the niche this sampler owns.