Cinematic Frame & Subtitle

Your model gives you a square or a 16:9 frame. What you want is the other thing - a still that looks like a frame grab from a 2.39:1 print with a yellow subtitle sitting on it. You can do that in Photoshop, or you can do it in the graph, deterministically, in about a fifth of a second, and keep the whole pipeline in ComfyUI. That's the job this node does.

It's the last creative step before you save. Cropping and typesetting aren't diffusion work - no model runs here, it's PIL and torch - which is exactly why you should do it in the graph instead of re-generating at some odd resolution and hoping the crop lands well.

How the frame gets chosen

aspect_ratio and lock_mode are deliberately identical to the pack's Aspect Ratio (1024) node, because the sizing math is literally the same calc_size() function imported from the same file. Same ratio list, same rule. So one ratio can drive both nodes and the frames come out pixel-identical - and if you feed this node the aspect_ratio string from Prompt Director (convert the widget to an input first), it's already legal, since that string is clamped to this exact list upstream.

lock_mode picks which side base_side pins. Short Side (recommended) is what you want 95% of the time: 21:9 stays 1024 high, 9:16 stays 1024 wide, so a scope frame and a vertical reel cost the same compute. At base_side 1024 that gives you 21:9 → 2384×1024, 16:9 → 1824×1024, 9:16 → 1024×1824. Both dimensions are always multiples of 16. Custom ratios work too - punch in 2.39:1 and you get 2448×1024.

Then fill_mode decides how your picture reaches that frame:

• Cinematic Crop (fill frame) - the default. Scales to fill and cuts the overflow off the centre. A 16:9 render dropped into 21:9 loses its top and bottom, never gains side bars. If your subject is centred this is invisible; if it isn't, check the result before you commit.
• Letterbox (black bars) - the only mode that shows the whole image, adding bars on whichever axis is left over.

The subtitle part

subtitle_top is set first and largest; subtitle_bottom sits under it at 82% of that size and is skipped entirely when empty. subtitle_size is a fraction of the frame's short side (0.045 default, ≈46px), which is why the line keeps the same optical weight in 21:9 and 9:16 instead of ballooning on a portrait frame. Long lines wrap at 86% of the picture width, per word for Latin and per character for CJK. subtitle_color gives you Yellow, White or Amber, and shadow (0.8 default) is a soft drop shadow scaled to the type, so the line survives a bright sky.

The optional group is where the polish lives. position = Bottom centres the line inside the lower black bar when Letterbox left room for one, and otherwise sits it just above the bottom edge of the picture; Center is the title-card placement. timecode takes 01:23:45:12 and burns a dashed box into the top-right with a TC prefix added for you - leave it empty and nothing is drawn. font_path empty falls back to DejaVu/Arial for Latin and Noto CJK/YaHei/PingFang when the line actually contains CJK characters.

One output: image. Wire it into your SaveImage, or into the pack's Watermark node.

Order matters, and the README is explicit about it: after the upscale, after Film Grain, before the watermark. Crop first so the subtitle lands inside the finished frame; watermark last so it sits on top rather than being cropped off the bottom edge.

Install

Search ComfyUI-FeiFei in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/aifeifei798/ComfyUI-FeiFei.git
# restart ComfyUI, nodes show up under the FeiFei category

No model downloads, no heavy dependencies. The pack's requirements.txt lists numpy, Pillow and openai>=1.40, and only the LLM nodes ever touch the last one. The repo also ships an example workflow at Workflow/ComfyUI-FeiFei.json - drag it onto the canvas and this node is already wired between the RTX upscale and the watermark at 2:1, Cinematic Crop, two English lines in Yellow.

Where people get burned

• Tofu boxes instead of Chinese/Japanese text. A missing font, not a node bug. Point font_path at a real CJK file, e.g. /usr/share/fonts/opentype/noto/NotoSansCJK-Regular.ttc.
• Heads getting cut. Cinematic Crop takes from the centre. Confirm the crop at your ratio, or switch to Letterbox and accept bars.
• The subtitle is baked in. It's pixels now - no layer, no undo. Duplicate the node with subtitle_top empty if you want a clean version of the same frame in the same run.
• The node isn't in the menu. Every node in this pack is imported inside its own try/except, so one bad node doesn't take the pack down. If everything else is there and this isn't, grep your console for [FeiFei] CinematicFrameSubtitle failed to load - the traceback is right under it.

For a batch of unrelated stills, skip this node. It's for the workflow where you're deliberately building a shot and every frame should match.